< prev index next >

src/hotspot/share/gc/shenandoah/shenandoahHeap.cpp

Print this page
@@ -1,8 +1,9 @@
  /*
   * Copyright (c) 2023, Oracle and/or its affiliates. All rights reserved.
   * Copyright (c) 2013, 2022, Red Hat, Inc. All rights reserved.
+  * Copyright Amazon.com Inc. or its affiliates. All Rights Reserved.
   * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   *
   * This code is free software; you can redistribute it and/or modify it
   * under the terms of the GNU General Public License version 2 only, as
   * published by the Free Software Foundation.

@@ -33,43 +34,56 @@
  #include "gc/shared/locationPrinter.inline.hpp"
  #include "gc/shared/memAllocator.hpp"
  #include "gc/shared/plab.hpp"
  #include "gc/shared/tlab_globals.hpp"
  
+ #include "gc/shenandoah/shenandoahAgeCensus.hpp"
+ #include "gc/shenandoah/heuristics/shenandoahOldHeuristics.hpp"
+ #include "gc/shenandoah/heuristics/shenandoahYoungHeuristics.hpp"
+ #include "gc/shenandoah/shenandoahAllocRequest.hpp"
  #include "gc/shenandoah/shenandoahBarrierSet.hpp"
+ #include "gc/shenandoah/shenandoahCardTable.hpp"
  #include "gc/shenandoah/shenandoahClosures.inline.hpp"
  #include "gc/shenandoah/shenandoahCollectionSet.hpp"
  #include "gc/shenandoah/shenandoahCollectorPolicy.hpp"
  #include "gc/shenandoah/shenandoahConcurrentMark.hpp"
  #include "gc/shenandoah/shenandoahMarkingContext.inline.hpp"
  #include "gc/shenandoah/shenandoahControlThread.hpp"
+ #include "gc/shenandoah/shenandoahRegulatorThread.hpp"
  #include "gc/shenandoah/shenandoahFreeSet.hpp"
+ #include "gc/shenandoah/shenandoahGlobalGeneration.hpp"
  #include "gc/shenandoah/shenandoahPhaseTimings.hpp"
  #include "gc/shenandoah/shenandoahHeap.inline.hpp"
  #include "gc/shenandoah/shenandoahHeapRegion.inline.hpp"
  #include "gc/shenandoah/shenandoahHeapRegionSet.hpp"
  #include "gc/shenandoah/shenandoahInitLogger.hpp"
  #include "gc/shenandoah/shenandoahMarkingContext.inline.hpp"
  #include "gc/shenandoah/shenandoahMemoryPool.hpp"
  #include "gc/shenandoah/shenandoahMetrics.hpp"
  #include "gc/shenandoah/shenandoahMonitoringSupport.hpp"
+ #include "gc/shenandoah/shenandoahOldGeneration.hpp"
  #include "gc/shenandoah/shenandoahOopClosures.inline.hpp"
  #include "gc/shenandoah/shenandoahPacer.inline.hpp"
  #include "gc/shenandoah/shenandoahPadding.hpp"
  #include "gc/shenandoah/shenandoahParallelCleaning.inline.hpp"
  #include "gc/shenandoah/shenandoahReferenceProcessor.hpp"
  #include "gc/shenandoah/shenandoahRootProcessor.inline.hpp"
+ #include "gc/shenandoah/shenandoahScanRemembered.inline.hpp"
  #include "gc/shenandoah/shenandoahSTWMark.hpp"
  #include "gc/shenandoah/shenandoahUtils.hpp"
  #include "gc/shenandoah/shenandoahVerifier.hpp"
  #include "gc/shenandoah/shenandoahCodeRoots.hpp"
  #include "gc/shenandoah/shenandoahVMOperations.hpp"
  #include "gc/shenandoah/shenandoahWorkGroup.hpp"
  #include "gc/shenandoah/shenandoahWorkerPolicy.hpp"
+ #include "gc/shenandoah/shenandoahYoungGeneration.hpp"
+ #include "gc/shenandoah/mode/shenandoahGenerationalMode.hpp"
  #include "gc/shenandoah/mode/shenandoahIUMode.hpp"
  #include "gc/shenandoah/mode/shenandoahPassiveMode.hpp"
  #include "gc/shenandoah/mode/shenandoahSATBMode.hpp"
+ #include "utilities/globalDefinitions.hpp"
+ 
  #if INCLUDE_JFR
  #include "gc/shenandoah/shenandoahJfrSupport.hpp"
  #endif
  
  #include "classfile/systemDictionary.hpp"

@@ -157,13 +171,10 @@
    _num_regions = ShenandoahHeapRegion::region_count();
    assert(_num_regions == (max_byte_size / reg_size_bytes),
           "Regions should cover entire heap exactly: " SIZE_FORMAT " != " SIZE_FORMAT "/" SIZE_FORMAT,
           _num_regions, max_byte_size, reg_size_bytes);
  
-   // Now we know the number of regions, initialize the heuristics.
-   initialize_heuristics();
- 
    size_t num_committed_regions = init_byte_size / reg_size_bytes;
    num_committed_regions = MIN2(num_committed_regions, _num_regions);
    assert(num_committed_regions <= _num_regions, "sanity");
    _initial_size = num_committed_regions * reg_size_bytes;
  

@@ -175,10 +186,13 @@
    // Default to max heap size.
    _soft_max_size = _num_regions * reg_size_bytes;
  
    _committed = _initial_size;
  
+   // Now we know the number of regions and heap sizes, initialize the heuristics.
+   initialize_heuristics_generations();
+ 
    size_t heap_page_size   = UseLargePages ? os::large_page_size() : os::vm_page_size();
    size_t bitmap_page_size = UseLargePages ? os::large_page_size() : os::vm_page_size();
    size_t region_page_size = UseLargePages ? os::large_page_size() : os::vm_page_size();
  
    //

@@ -213,10 +227,38 @@
    if (!_heap_region_special) {
      os::commit_memory_or_exit(sh_rs.base(), _initial_size, heap_alignment, false,
                                "Cannot commit heap memory");
    }
  
+   BarrierSet::set_barrier_set(new ShenandoahBarrierSet(this, _heap_region));
+ 
+   //
+   // After reserving the Java heap, create the card table, barriers, and workers, in dependency order
+   //
+   if (mode()->is_generational()) {
+     ShenandoahDirectCardMarkRememberedSet *rs;
+     ShenandoahCardTable* card_table = ShenandoahBarrierSet::barrier_set()->card_table();
+     size_t card_count = card_table->cards_required(heap_rs.size() / HeapWordSize);
+     rs = new ShenandoahDirectCardMarkRememberedSet(ShenandoahBarrierSet::barrier_set()->card_table(), card_count);
+     _card_scan = new ShenandoahScanRemembered<ShenandoahDirectCardMarkRememberedSet>(rs);
+ 
+     // Age census structure
+     _age_census = new ShenandoahAgeCensus();
+   }
+ 
+   _workers = new ShenandoahWorkerThreads("Shenandoah GC Threads", _max_workers);
+   if (_workers == nullptr) {
+     vm_exit_during_initialization("Failed necessary allocation.");
+   } else {
+     _workers->initialize_workers();
+   }
+ 
+   if (ParallelGCThreads > 1) {
+     _safepoint_workers = new ShenandoahWorkerThreads("Safepoint Cleanup Thread", ParallelGCThreads);
+     _safepoint_workers->initialize_workers();
+   }
+ 
    //
    // Reserve and commit memory for bitmap(s)
    //
  
    size_t bitmap_size_orig = ShenandoahMarkBitMap::compute_size(heap_rs.size());

@@ -260,11 +302,11 @@
    if (!_bitmap_region_special) {
      os::commit_memory_or_exit((char *) _bitmap_region.start(), bitmap_init_commit, bitmap_page_size, false,
                                "Cannot commit bitmap memory");
    }
  
-   _marking_context = new ShenandoahMarkingContext(_heap_region, _bitmap_region, _num_regions, _max_workers);
+   _marking_context = new ShenandoahMarkingContext(_heap_region, _bitmap_region, _num_regions);
  
    if (ShenandoahVerify) {
      ReservedSpace verify_bitmap(_bitmap_size, bitmap_page_size);
      os::trace_page_sizes_for_requested_size("Verify Bitmap",
                                              bitmap_size_orig, bitmap_page_size,

@@ -351,15 +393,17 @@
                                              cset_rs.base(),
                                              cset_rs.size(), cset_rs.page_size());
    }
  
    _regions = NEW_C_HEAP_ARRAY(ShenandoahHeapRegion*, _num_regions, mtGC);
+   _affiliations = NEW_C_HEAP_ARRAY(uint8_t, _num_regions, mtGC);
    _free_set = new ShenandoahFreeSet(this, _num_regions);
  
    {
      ShenandoahHeapLocker locker(lock());
  
+ 
      for (size_t i = 0; i < _num_regions; i++) {
        HeapWord* start = (HeapWord*)sh_rs.base() + ShenandoahHeapRegion::region_size_words() * i;
        bool is_committed = i < num_committed_regions;
        void* loc = region_storage.base() + i * region_align;
  

@@ -367,16 +411,21 @@
        assert(is_aligned(r, SHENANDOAH_CACHE_LINE_SIZE), "Sanity");
  
        _marking_context->initialize_top_at_mark_start(r);
        _regions[i] = r;
        assert(!collection_set()->is_in(i), "New region should not be in collection set");
+ 
+       _affiliations[i] = ShenandoahAffiliation::FREE;
      }
  
      // Initialize to complete
      _marking_context->mark_complete();
+     size_t young_cset_regions, old_cset_regions;
  
-     _free_set->rebuild();
+     // We are initializing free set.  We ignore cset region tallies.
+     _free_set->prepare_to_rebuild(young_cset_regions, old_cset_regions);
+     _free_set->rebuild(young_cset_regions, old_cset_regions);
    }
  
    if (AlwaysPreTouch) {
      // For NUMA, it is important to pre-touch the storage under bitmaps with worker threads,
      // before initialize() below zeroes it with initializing thread. For any given region,

@@ -434,24 +483,61 @@
    } else {
      _pacer = nullptr;
    }
  
    _control_thread = new ShenandoahControlThread();
+   _regulator_thread = new ShenandoahRegulatorThread(_control_thread);
  
-   ShenandoahInitLogger::print();
+   print_init_logger();
  
    return JNI_OK;
  }
  
- void ShenandoahHeap::initialize_mode() {
+ void ShenandoahHeap::print_init_logger() const {
+   ShenandoahInitLogger::print();
+ }
+ 
+ size_t ShenandoahHeap::max_size_for(ShenandoahGeneration* generation) const {
+   switch (generation->type()) {
+     case YOUNG:
+       return _generation_sizer.max_young_size();
+     case OLD:
+       return max_capacity() - _generation_sizer.min_young_size();
+     case GLOBAL_GEN:
+     case GLOBAL_NON_GEN:
+       return max_capacity();
+     default:
+       ShouldNotReachHere();
+       return 0;
+   }
+ }
+ 
+ size_t ShenandoahHeap::min_size_for(ShenandoahGeneration* generation) const {
+   switch (generation->type()) {
+     case YOUNG:
+       return _generation_sizer.min_young_size();
+     case OLD:
+       return max_capacity() - _generation_sizer.max_young_size();
+     case GLOBAL_GEN:
+     case GLOBAL_NON_GEN:
+       return min_capacity();
+     default:
+       ShouldNotReachHere();
+       return 0;
+   }
+ }
+ 
+ void ShenandoahHeap::initialize_heuristics_generations() {
    if (ShenandoahGCMode != nullptr) {
      if (strcmp(ShenandoahGCMode, "satb") == 0) {
        _gc_mode = new ShenandoahSATBMode();
      } else if (strcmp(ShenandoahGCMode, "iu") == 0) {
        _gc_mode = new ShenandoahIUMode();
      } else if (strcmp(ShenandoahGCMode, "passive") == 0) {
        _gc_mode = new ShenandoahPassiveMode();
+     } else if (strcmp(ShenandoahGCMode, "generational") == 0) {
+       _gc_mode = new ShenandoahGenerationalMode();
      } else {
        vm_exit_during_initialization("Unknown -XX:ShenandoahGCMode option");
      }
    } else {
      vm_exit_during_initialization("Unknown -XX:ShenandoahGCMode option (null)");

@@ -465,123 +551,102 @@
    if (_gc_mode->is_experimental() && !UnlockExperimentalVMOptions) {
      vm_exit_during_initialization(
              err_msg("GC mode \"%s\" is experimental, and must be enabled via -XX:+UnlockExperimentalVMOptions.",
                      _gc_mode->name()));
    }
- }
  
- void ShenandoahHeap::initialize_heuristics() {
-   assert(_gc_mode != nullptr, "Must be initialized");
-   _heuristics = _gc_mode->initialize_heuristics();
+   // Max capacity is the maximum _allowed_ capacity. That is, the maximum allowed capacity
+   // for old would be total heap - minimum capacity of young. This means the sum of the maximum
+   // allowed for old and young could exceed the total heap size. It remains the case that the
+   // _actual_ capacity of young + old = total.
+   _generation_sizer.heap_size_changed(max_capacity());
+   size_t initial_capacity_young = _generation_sizer.max_young_size();
+   size_t max_capacity_young = _generation_sizer.max_young_size();
+   size_t initial_capacity_old = max_capacity() - max_capacity_young;
+   size_t max_capacity_old = max_capacity() - initial_capacity_young;
  
-   if (_heuristics->is_diagnostic() && !UnlockDiagnosticVMOptions) {
-     vm_exit_during_initialization(
-             err_msg("Heuristics \"%s\" is diagnostic, and must be enabled via -XX:+UnlockDiagnosticVMOptions.",
-                     _heuristics->name()));
-   }
-   if (_heuristics->is_experimental() && !UnlockExperimentalVMOptions) {
-     vm_exit_during_initialization(
-             err_msg("Heuristics \"%s\" is experimental, and must be enabled via -XX:+UnlockExperimentalVMOptions.",
-                     _heuristics->name()));
+   _young_generation = new ShenandoahYoungGeneration(_max_workers, max_capacity_young, initial_capacity_young);
+   _old_generation = new ShenandoahOldGeneration(_max_workers, max_capacity_old, initial_capacity_old);
+   _global_generation = new ShenandoahGlobalGeneration(_gc_mode->is_generational(), _max_workers, max_capacity(), max_capacity());
+   _global_generation->initialize_heuristics(_gc_mode);
+   if (mode()->is_generational()) {
+     _young_generation->initialize_heuristics(_gc_mode);
+     _old_generation->initialize_heuristics(_gc_mode);
    }
+   _evac_tracker = new ShenandoahEvacuationTracker(mode()->is_generational());
  }
  
  #ifdef _MSC_VER
  #pragma warning( push )
  #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
  #endif
  
  ShenandoahHeap::ShenandoahHeap(ShenandoahCollectorPolicy* policy) :
    CollectedHeap(),
+   _gc_generation(nullptr),
+   _prepare_for_old_mark(false),
    _initial_size(0),
-   _used(0),
+   _promotion_potential(0),
    _committed(0),
-   _bytes_allocated_since_gc_start(0),
-   _max_workers(MAX2(ConcGCThreads, ParallelGCThreads)),
+   _max_workers(MAX3(ConcGCThreads, ParallelGCThreads, 1U)),
    _workers(nullptr),
    _safepoint_workers(nullptr),
    _heap_region_special(false),
    _num_regions(0),
    _regions(nullptr),
+   _affiliations(nullptr),
    _update_refs_iterator(this),
+   _promoted_reserve(0),
+   _old_evac_reserve(0),
+   _old_evac_expended(0),
+   _young_evac_reserve(0),
+   _captured_old_usage(0),
+   _previous_promotion(0),
+   _upgraded_to_full(false),
+   _age_census(nullptr),
+   _has_evacuation_reserve_quantities(false),
+   _cancel_requested_time(0),
+   _young_generation(nullptr),
+   _global_generation(nullptr),
+   _old_generation(nullptr),
    _control_thread(nullptr),
+   _regulator_thread(nullptr),
    _shenandoah_policy(policy),
-   _gc_mode(nullptr),
-   _heuristics(nullptr),
    _free_set(nullptr),
    _pacer(nullptr),
    _verifier(nullptr),
    _phase_timings(nullptr),
+   _evac_tracker(nullptr),
+   _mmu_tracker(),
+   _generation_sizer(&_mmu_tracker),
    _monitoring_support(nullptr),
    _memory_pool(nullptr),
+   _young_gen_memory_pool(nullptr),
+   _old_gen_memory_pool(nullptr),
    _stw_memory_manager("Shenandoah Pauses"),
    _cycle_memory_manager("Shenandoah Cycles"),
    _gc_timer(new ConcurrentGCTimer()),
    _soft_ref_policy(),
    _log_min_obj_alignment_in_bytes(LogMinObjAlignmentInBytes),
-   _ref_processor(new ShenandoahReferenceProcessor(MAX2(_max_workers, 1U))),
+   _old_regions_surplus(0),
+   _old_regions_deficit(0),
    _marking_context(nullptr),
    _bitmap_size(0),
    _bitmap_regions_per_slice(0),
    _bitmap_bytes_per_slice(0),
    _bitmap_region_special(false),
    _aux_bitmap_region_special(false),
    _liveness_cache(nullptr),
-   _collection_set(nullptr)
+   _collection_set(nullptr),
+   _card_scan(nullptr)
  {
-   // Initialize GC mode early, so we can adjust barrier support
-   initialize_mode();
-   BarrierSet::set_barrier_set(new ShenandoahBarrierSet(this));
- 
-   _max_workers = MAX2(_max_workers, 1U);
-   _workers = new ShenandoahWorkerThreads("Shenandoah GC Threads", _max_workers);
-   if (_workers == nullptr) {
-     vm_exit_during_initialization("Failed necessary allocation.");
-   } else {
-     _workers->initialize_workers();
-   }
- 
-   if (ParallelGCThreads > 1) {
-     _safepoint_workers = new ShenandoahWorkerThreads("Safepoint Cleanup Thread",
-                                                 ParallelGCThreads);
-     _safepoint_workers->initialize_workers();
-   }
  }
  
  #ifdef _MSC_VER
  #pragma warning( pop )
  #endif
  
- class ShenandoahResetBitmapTask : public WorkerTask {
- private:
-   ShenandoahRegionIterator _regions;
- 
- public:
-   ShenandoahResetBitmapTask() :
-     WorkerTask("Shenandoah Reset Bitmap") {}
- 
-   void work(uint worker_id) {
-     ShenandoahHeapRegion* region = _regions.next();
-     ShenandoahHeap* heap = ShenandoahHeap::heap();
-     ShenandoahMarkingContext* const ctx = heap->marking_context();
-     while (region != nullptr) {
-       if (heap->is_bitmap_slice_committed(region)) {
-         ctx->clear_bitmap(region);
-       }
-       region = _regions.next();
-     }
-   }
- };
- 
- void ShenandoahHeap::reset_mark_bitmap() {
-   assert_gc_workers(_workers->active_workers());
-   mark_incomplete_marking_context();
- 
-   ShenandoahResetBitmapTask task;
-   _workers->run_task(&task);
- }
- 
  void ShenandoahHeap::print_on(outputStream* st) const {
    st->print_cr("Shenandoah Heap");
    st->print_cr(" " SIZE_FORMAT "%s max, " SIZE_FORMAT "%s soft max, " SIZE_FORMAT "%s committed, " SIZE_FORMAT "%s used",
                 byte_size_in_proper_unit(max_capacity()), proper_unit_for_byte_size(max_capacity()),
                 byte_size_in_proper_unit(soft_max_capacity()), proper_unit_for_byte_size(soft_max_capacity()),

@@ -592,11 +657,12 @@
                 byte_size_in_proper_unit(ShenandoahHeapRegion::region_size_bytes()),
                 proper_unit_for_byte_size(ShenandoahHeapRegion::region_size_bytes()));
  
    st->print("Status: ");
    if (has_forwarded_objects())                 st->print("has forwarded objects, ");
-   if (is_concurrent_mark_in_progress())        st->print("marking, ");
+   if (is_concurrent_old_mark_in_progress())    st->print("old marking, ");
+   if (is_concurrent_young_mark_in_progress())  st->print("young marking, ");
    if (is_evacuation_in_progress())             st->print("evacuating, ");
    if (is_update_refs_in_progress())            st->print("updating refs, ");
    if (is_degenerated_gc_in_progress())         st->print("degenerated gc, ");
    if (is_full_gc_in_progress())                st->print("full gc, ");
    if (is_full_gc_move_in_progress())           st->print("full gc move, ");

@@ -643,10 +709,12 @@
    }
  };
  
  void ShenandoahHeap::post_initialize() {
    CollectedHeap::post_initialize();
+   _mmu_tracker.initialize();
+ 
    MutexLocker ml(Threads_lock);
  
    ShenandoahInitWorkerGCLABClosure init_gclabs;
    _workers->threads_do(&init_gclabs);
  

@@ -656,64 +724,133 @@
    if (_safepoint_workers != nullptr) {
      _safepoint_workers->threads_do(&init_gclabs);
      _safepoint_workers->set_initialize_gclab();
    }
  
-   _heuristics->initialize();
- 
    JFR_ONLY(ShenandoahJFRSupport::register_jfr_type_serializers());
  }
  
+ ShenandoahHeuristics* ShenandoahHeap::heuristics() {
+   return _global_generation->heuristics();
+ }
+ 
+ ShenandoahOldHeuristics* ShenandoahHeap::old_heuristics() {
+   return (ShenandoahOldHeuristics*) _old_generation->heuristics();
+ }
+ 
+ ShenandoahYoungHeuristics* ShenandoahHeap::young_heuristics() {
+   return (ShenandoahYoungHeuristics*) _young_generation->heuristics();
+ }
+ 
+ bool ShenandoahHeap::doing_mixed_evacuations() {
+   return _old_generation->state() == ShenandoahOldGeneration::WAITING_FOR_EVAC;
+ }
+ 
+ bool ShenandoahHeap::is_old_bitmap_stable() const {
+   return _old_generation->is_mark_complete();
+ }
+ 
+ bool ShenandoahHeap::is_gc_generation_young() const {
+   return _gc_generation != nullptr && _gc_generation->is_young();
+ }
+ 
  size_t ShenandoahHeap::used() const {
-   return Atomic::load(&_used);
+   return global_generation()->used();
  }
  
  size_t ShenandoahHeap::committed() const {
    return Atomic::load(&_committed);
  }
  
- size_t ShenandoahHeap::available() const {
-   return free_set()->available();
- }
- 
  void ShenandoahHeap::increase_committed(size_t bytes) {
    shenandoah_assert_heaplocked_or_safepoint();
    _committed += bytes;
  }
  
  void ShenandoahHeap::decrease_committed(size_t bytes) {
    shenandoah_assert_heaplocked_or_safepoint();
    _committed -= bytes;
  }
  
- void ShenandoahHeap::increase_used(size_t bytes) {
-   Atomic::add(&_used, bytes, memory_order_relaxed);
+ // For tracking usage based on allocations, it should be the case that:
+ // * The sum of regions::used == heap::used
+ // * The sum of a generation's regions::used == generation::used
+ // * The sum of a generation's humongous regions::free == generation::humongous_waste
+ // These invariants are checked by the verifier on GC safepoints.
+ //
+ // Additional notes:
+ // * When a mutator's allocation request causes a region to be retired, the
+ //   free memory left in that region is considered waste. It does not contribute
+ //   to the usage, but it _does_ contribute to allocation rate.
+ // * The bottom of a PLAB must be aligned on card size. In some cases this will
+ //   require padding in front of the PLAB (a filler object). Because this padding
+ //   is included in the region's used memory we include the padding in the usage
+ //   accounting as waste.
+ // * Mutator allocations are used to compute an allocation rate. They are also
+ //   sent to the Pacer for those purposes.
+ // * There are three sources of waste:
+ //  1. The padding used to align a PLAB on card size
+ //  2. Region's free is less than minimum TLAB size and is retired
+ //  3. The unused portion of memory in the last region of a humongous object
+ void ShenandoahHeap::increase_used(const ShenandoahAllocRequest& req) {
+   size_t actual_bytes = req.actual_size() * HeapWordSize;
+   size_t wasted_bytes = req.waste() * HeapWordSize;
+   ShenandoahGeneration* generation = generation_for(req.affiliation());
+ 
+   if (req.is_gc_alloc()) {
+     assert(wasted_bytes == 0 || req.type() == ShenandoahAllocRequest::_alloc_plab, "Only PLABs have waste");
+     increase_used(generation, actual_bytes + wasted_bytes);
+   } else {
+     assert(req.is_mutator_alloc(), "Expected mutator alloc here");
+     // padding and actual size both count towards allocation counter
+     generation->increase_allocated(actual_bytes + wasted_bytes);
+ 
+     // only actual size counts toward usage for mutator allocations
+     increase_used(generation, actual_bytes);
+ 
+     // notify pacer of both actual size and waste
+     notify_mutator_alloc_words(req.actual_size(), req.waste());
+ 
+     if (wasted_bytes > 0 && req.actual_size() > ShenandoahHeapRegion::humongous_threshold_words()) {
+       increase_humongous_waste(generation,wasted_bytes);
+     }
+   }
  }
  
- void ShenandoahHeap::set_used(size_t bytes) {
-   Atomic::store(&_used, bytes);
+ void ShenandoahHeap::increase_humongous_waste(ShenandoahGeneration* generation, size_t bytes) {
+   generation->increase_humongous_waste(bytes);
+   if (!generation->is_global()) {
+     global_generation()->increase_humongous_waste(bytes);
+   }
  }
  
- void ShenandoahHeap::decrease_used(size_t bytes) {
-   assert(used() >= bytes, "never decrease heap size by more than we've left");
-   Atomic::sub(&_used, bytes, memory_order_relaxed);
+ void ShenandoahHeap::decrease_humongous_waste(ShenandoahGeneration* generation, size_t bytes) {
+   generation->decrease_humongous_waste(bytes);
+   if (!generation->is_global()) {
+     global_generation()->decrease_humongous_waste(bytes);
+   }
  }
  
- void ShenandoahHeap::increase_allocated(size_t bytes) {
-   Atomic::add(&_bytes_allocated_since_gc_start, bytes, memory_order_relaxed);
+ void ShenandoahHeap::increase_used(ShenandoahGeneration* generation, size_t bytes) {
+   generation->increase_used(bytes);
+   if (!generation->is_global()) {
+     global_generation()->increase_used(bytes);
+   }
  }
  
- void ShenandoahHeap::notify_mutator_alloc_words(size_t words, bool waste) {
-   size_t bytes = words * HeapWordSize;
-   if (!waste) {
-     increase_used(bytes);
+ void ShenandoahHeap::decrease_used(ShenandoahGeneration* generation, size_t bytes) {
+   generation->decrease_used(bytes);
+   if (!generation->is_global()) {
+     global_generation()->decrease_used(bytes);
    }
-   increase_allocated(bytes);
+ }
+ 
+ void ShenandoahHeap::notify_mutator_alloc_words(size_t words, size_t waste) {
    if (ShenandoahPacing) {
      control_thread()->pacing_notify_alloc(words);
-     if (waste) {
-       pacer()->claim_for_alloc(words, true);
+     if (waste > 0) {
+       pacer()->claim_for_alloc(waste, true);
      }
    }
  }
  
  size_t ShenandoahHeap::capacity() const {

@@ -745,16 +882,10 @@
  
  size_t ShenandoahHeap::initial_capacity() const {
    return _initial_size;
  }
  
- bool ShenandoahHeap::is_in(const void* p) const {
-   HeapWord* heap_base = (HeapWord*) base();
-   HeapWord* last_region_end = heap_base + ShenandoahHeapRegion::region_size_words() * num_regions();
-   return p >= heap_base && p < last_region_end;
- }
- 
  void ShenandoahHeap::op_uncommit(double shrink_before, size_t shrink_until) {
    assert (ShenandoahUncommit, "should be enabled");
  
    // Application allocates from the beginning of the heap, and GC allocates at
    // the end of it. It is more efficient to uncommit from the end, so that applications

@@ -778,19 +909,92 @@
      SpinPause(); // allow allocators to take the lock
    }
  
    if (count > 0) {
      control_thread()->notify_heap_changed();
+     regulator_thread()->notify_heap_changed();
+   }
+ }
+ 
+ void ShenandoahHeap::handle_old_evacuation(HeapWord* obj, size_t words, bool promotion) {
+   // Only register the copy of the object that won the evacuation race.
+   card_scan()->register_object_without_lock(obj);
+ 
+   // Mark the entire range of the evacuated object as dirty.  At next remembered set scan,
+   // we will clear dirty bits that do not hold interesting pointers.  It's more efficient to
+   // do this in batch, in a background GC thread than to try to carefully dirty only cards
+   // that hold interesting pointers right now.
+   card_scan()->mark_range_as_dirty(obj, words);
+ 
+   if (promotion) {
+     // This evacuation was a promotion, track this as allocation against old gen
+     old_generation()->increase_allocated(words * HeapWordSize);
+   }
+ }
+ 
+ void ShenandoahHeap::handle_old_evacuation_failure() {
+   if (_old_gen_oom_evac.try_set()) {
+     log_info(gc)("Old gen evac failure.");
+   }
+ }
+ 
+ void ShenandoahHeap::report_promotion_failure(Thread* thread, size_t size) {
+   // We squelch excessive reports to reduce noise in logs.
+   const size_t MaxReportsPerEpoch = 4;
+   static size_t last_report_epoch = 0;
+   static size_t epoch_report_count = 0;
+ 
+   size_t promotion_reserve;
+   size_t promotion_expended;
+ 
+   size_t gc_id = control_thread()->get_gc_id();
+ 
+   if ((gc_id != last_report_epoch) || (epoch_report_count++ < MaxReportsPerEpoch)) {
+     {
+       // Promotion failures should be very rare.  Invest in providing useful diagnostic info.
+       ShenandoahHeapLocker locker(lock());
+       promotion_reserve = get_promoted_reserve();
+       promotion_expended = get_promoted_expended();
+     }
+     PLAB* plab = ShenandoahThreadLocalData::plab(thread);
+     size_t words_remaining = (plab == nullptr)? 0: plab->words_remaining();
+     const char* promote_enabled = ShenandoahThreadLocalData::allow_plab_promotions(thread)? "enabled": "disabled";
+     ShenandoahGeneration* old_gen = old_generation();
+     size_t old_capacity = old_gen->max_capacity();
+     size_t old_usage = old_gen->used();
+     size_t old_free_regions = old_gen->free_unaffiliated_regions();
+ 
+     log_info(gc, ergo)("Promotion failed, size " SIZE_FORMAT ", has plab? %s, PLAB remaining: " SIZE_FORMAT
+                        ", plab promotions %s, promotion reserve: " SIZE_FORMAT ", promotion expended: " SIZE_FORMAT
+                        ", old capacity: " SIZE_FORMAT ", old_used: " SIZE_FORMAT ", old unaffiliated regions: " SIZE_FORMAT,
+                        size * HeapWordSize, plab == nullptr? "no": "yes",
+                        words_remaining * HeapWordSize, promote_enabled, promotion_reserve, promotion_expended,
+                        old_capacity, old_usage, old_free_regions);
+ 
+     if ((gc_id == last_report_epoch) && (epoch_report_count >= MaxReportsPerEpoch)) {
+       log_info(gc, ergo)("Squelching additional promotion failure reports for current epoch");
+     } else if (gc_id != last_report_epoch) {
+       last_report_epoch = gc_id;;
+       epoch_report_count = 1;
+     }
    }
  }
  
  HeapWord* ShenandoahHeap::allocate_from_gclab_slow(Thread* thread, size_t size) {
    // New object should fit the GCLAB size
    size_t min_size = MAX2(size, PLAB::min_size());
  
    // Figure out size of new GCLAB, looking back at heuristics. Expand aggressively.
    size_t new_size = ShenandoahThreadLocalData::gclab_size(thread) * 2;
+ 
+   // Limit growth of GCLABs to ShenandoahMaxEvacLABRatio * the minimum size.  This enables more equitable distribution of
+   // available evacuation buidget between the many threads that are coordinating in the evacuation effort.
+   if (ShenandoahMaxEvacLABRatio > 0) {
+     log_debug(gc, free)("Allocate new gclab: " SIZE_FORMAT ", " SIZE_FORMAT, new_size, PLAB::min_size() * ShenandoahMaxEvacLABRatio);
+     new_size = MIN2(new_size, PLAB::min_size() * ShenandoahMaxEvacLABRatio);
+   }
+ 
    new_size = MIN2(new_size, PLAB::max_size());
    new_size = MAX2(new_size, PLAB::min_size());
  
    // Record new heuristic value even if we take any shortcut. This captures
    // the case when moderately-sized objects always take a shortcut. At some point,

@@ -798,10 +1002,11 @@
    ShenandoahThreadLocalData::set_gclab_size(thread, new_size);
  
    if (new_size < size) {
      // New size still does not fit the object. Fall back to shared allocation.
      // This avoids retiring perfectly good GCLABs, when we encounter a large object.
+     log_debug(gc, free)("New gclab size (" SIZE_FORMAT ") is too small for " SIZE_FORMAT, new_size, size);
      return nullptr;
    }
  
    // Retire current GCLAB, and allocate a new one.
    PLAB* gclab = ShenandoahThreadLocalData::gclab(thread);

@@ -830,15 +1035,252 @@
    }
    gclab->set_buf(gclab_buf, actual_size);
    return gclab->allocate(size);
  }
  
+ // Establish a new PLAB and allocate size HeapWords within it.
+ HeapWord* ShenandoahHeap::allocate_from_plab_slow(Thread* thread, size_t size, bool is_promotion) {
+   // New object should fit the PLAB size
+   size_t min_size = MAX2(size, PLAB::min_size());
+ 
+   // Figure out size of new PLAB, looking back at heuristics. Expand aggressively.
+   size_t cur_size = ShenandoahThreadLocalData::plab_size(thread);
+   if (cur_size == 0) {
+     cur_size = PLAB::min_size();
+   }
+   size_t future_size = cur_size * 2;
+   // Limit growth of PLABs to ShenandoahMaxEvacLABRatio * the minimum size.  This enables more equitable distribution of
+   // available evacuation buidget between the many threads that are coordinating in the evacuation effort.
+   if (ShenandoahMaxEvacLABRatio > 0) {
+     future_size = MIN2(future_size, PLAB::min_size() * ShenandoahMaxEvacLABRatio);
+   }
+   future_size = MIN2(future_size, PLAB::max_size());
+   future_size = MAX2(future_size, PLAB::min_size());
+ 
+   size_t unalignment = future_size % CardTable::card_size_in_words();
+   if (unalignment != 0) {
+     future_size = future_size - unalignment + CardTable::card_size_in_words();
+   }
+ 
+   // Record new heuristic value even if we take any shortcut. This captures
+   // the case when moderately-sized objects always take a shortcut. At some point,
+   // heuristics should catch up with them.  Note that the requested cur_size may
+   // not be honored, but we remember that this is the preferred size.
+   ShenandoahThreadLocalData::set_plab_size(thread, future_size);
+   if (cur_size < size) {
+     // The PLAB to be allocated is still not large enough to hold the object. Fall back to shared allocation.
+     // This avoids retiring perfectly good PLABs in order to represent a single large object allocation.
+     return nullptr;
+   }
+ 
+   // Retire current PLAB, and allocate a new one.
+   PLAB* plab = ShenandoahThreadLocalData::plab(thread);
+   if (plab->words_remaining() < PLAB::min_size()) {
+     // Retire current PLAB, and allocate a new one.
+     // CAUTION: retire_plab may register the remnant filler object with the remembered set scanner without a lock.  This
+     // is safe iff it is assured that each PLAB is a whole-number multiple of card-mark memory size and each PLAB is
+     // aligned with the start of a card's memory range.
+     retire_plab(plab, thread);
+ 
+     size_t actual_size = 0;
+     // allocate_new_plab resets plab_evacuated and plab_promoted and disables promotions if old-gen available is
+     // less than the remaining evacuation need.  It also adjusts plab_preallocated and expend_promoted if appropriate.
+     HeapWord* plab_buf = allocate_new_plab(min_size, cur_size, &actual_size);
+     if (plab_buf == nullptr) {
+       if (min_size == PLAB::min_size()) {
+         // Disable plab promotions for this thread because we cannot even allocate a plab of minimal size.  This allows us
+         // to fail faster on subsequent promotion attempts.
+         ShenandoahThreadLocalData::disable_plab_promotions(thread);
+       }
+       return NULL;
+     } else {
+       ShenandoahThreadLocalData::enable_plab_retries(thread);
+     }
+     assert (size <= actual_size, "allocation should fit");
+     if (ZeroTLAB) {
+       // ..and clear it.
+       Copy::zero_to_words(plab_buf, actual_size);
+     } else {
+       // ...and zap just allocated object.
+ #ifdef ASSERT
+       // Skip mangling the space corresponding to the object header to
+       // ensure that the returned space is not considered parsable by
+       // any concurrent GC thread.
+       size_t hdr_size = oopDesc::header_size();
+       Copy::fill_to_words(plab_buf + hdr_size, actual_size - hdr_size, badHeapWordVal);
+ #endif // ASSERT
+     }
+     plab->set_buf(plab_buf, actual_size);
+     if (is_promotion && !ShenandoahThreadLocalData::allow_plab_promotions(thread)) {
+       return nullptr;
+     }
+     return plab->allocate(size);
+   } else {
+     // If there's still at least min_size() words available within the current plab, don't retire it.  Let's gnaw
+     // away on this plab as long as we can.  Meanwhile, return nullptr to force this particular allocation request
+     // to be satisfied with a shared allocation.  By packing more promotions into the previously allocated PLAB, we
+     // reduce the likelihood of evacuation failures, and we we reduce the need for downsizing our PLABs.
+     return nullptr;
+   }
+ }
+ 
+ // TODO: It is probably most efficient to register all objects (both promotions and evacuations) that were allocated within
+ // this plab at the time we retire the plab.  A tight registration loop will run within both code and data caches.  This change
+ // would allow smaller and faster in-line implementation of alloc_from_plab().  Since plabs are aligned on card-table boundaries,
+ // this object registration loop can be performed without acquiring a lock.
+ void ShenandoahHeap::retire_plab(PLAB* plab, Thread* thread) {
+   // We don't enforce limits on plab_evacuated.  We let it consume all available old-gen memory in order to reduce
+   // probability of an evacuation failure.  We do enforce limits on promotion, to make sure that excessive promotion
+   // does not result in an old-gen evacuation failure.  Note that a failed promotion is relatively harmless.  Any
+   // object that fails to promote in the current cycle will be eligible for promotion in a subsequent cycle.
+ 
+   // When the plab was instantiated, its entirety was treated as if the entire buffer was going to be dedicated to
+   // promotions.  Now that we are retiring the buffer, we adjust for the reality that the plab is not entirely promotions.
+   //  1. Some of the plab may have been dedicated to evacuations.
+   //  2. Some of the plab may have been abandoned due to waste (at the end of the plab).
+   size_t not_promoted =
+     ShenandoahThreadLocalData::get_plab_preallocated_promoted(thread) - ShenandoahThreadLocalData::get_plab_promoted(thread);
+   ShenandoahThreadLocalData::reset_plab_promoted(thread);
+   ShenandoahThreadLocalData::reset_plab_evacuated(thread);
+   ShenandoahThreadLocalData::set_plab_preallocated_promoted(thread, 0);
+   if (not_promoted > 0) {
+     unexpend_promoted(not_promoted);
+   }
+   size_t waste = plab->waste();
+   HeapWord* top = plab->top();
+   plab->retire();
+   if (top != nullptr && plab->waste() > waste && is_in_old(top)) {
+     // If retiring the plab created a filler object, then we
+     // need to register it with our card scanner so it can
+     // safely walk the region backing the plab.
+     log_debug(gc)("retire_plab() is registering remnant of size " SIZE_FORMAT " at " PTR_FORMAT,
+                   plab->waste() - waste, p2i(top));
+     card_scan()->register_object_without_lock(top);
+   }
+ }
+ 
+ void ShenandoahHeap::retire_plab(PLAB* plab) {
+   Thread* thread = Thread::current();
+   retire_plab(plab, thread);
+ }
+ 
+ void ShenandoahHeap::cancel_old_gc() {
+   shenandoah_assert_safepoint();
+   assert(_old_generation != nullptr, "Should only have mixed collections in generation mode.");
+   log_info(gc)("Terminating old gc cycle.");
+ 
+   // Stop marking
+   old_generation()->cancel_marking();
+   // Stop coalescing undead objects
+   set_prepare_for_old_mark_in_progress(false);
+   // Stop tracking old regions
+   old_heuristics()->abandon_collection_candidates();
+   // Remove old generation access to young generation mark queues
+   young_generation()->set_old_gen_task_queues(nullptr);
+   // Transition to IDLE now.
+   _old_generation->transition_to(ShenandoahOldGeneration::IDLE);
+ }
+ 
+ bool ShenandoahHeap::is_old_gc_active() {
+   return _old_generation->state() != ShenandoahOldGeneration::IDLE;
+ }
+ 
+ // xfer_limit is the maximum we're able to transfer from young to old
+ void ShenandoahHeap::adjust_generation_sizes_for_next_cycle(
+   size_t xfer_limit, size_t young_cset_regions, size_t old_cset_regions) {
+ 
+   // Make sure old-generation is large enough, but no larger, than is necessary to hold mixed evacuations
+   // and promotions if we anticipate either.
+   size_t region_size_bytes = ShenandoahHeapRegion::region_size_bytes();
+   size_t promo_load = get_promotion_potential();
+   // The free set will reserve this amount of memory to hold young evacuations
+   size_t young_reserve = (young_generation()->max_capacity() * ShenandoahEvacReserve) / 100;
+   size_t old_reserve = 0;
+   size_t mixed_candidates = old_heuristics()->unprocessed_old_collection_candidates();
+   bool doing_mixed = (mixed_candidates > 0);
+   bool doing_promotions = promo_load > 0;
+ 
+   // round down
+   size_t max_old_region_xfer = xfer_limit / region_size_bytes;
+ 
+   // We can limit the reserve to the size of anticipated promotions
+   size_t max_old_reserve = young_reserve * ShenandoahOldEvacRatioPercent / (100 - ShenandoahOldEvacRatioPercent);
+   // Here's the algebra:
+   //  TotalEvacuation = OldEvacuation + YoungEvacuation
+   //  OldEvacuation = TotalEvacuation*(ShenandoahOldEvacRatioPercent/100)
+   //  OldEvacuation = YoungEvacuation * (ShenandoahOldEvacRatioPercent/100)/(1 - ShenandoahOldEvacRatioPercent/100)
+   //  OldEvacuation = YoungEvacuation * ShenandoahOldEvacRatioPercent/(100 - ShenandoahOldEvacRatioPercent)
+ 
+   size_t reserve_for_mixed, reserve_for_promo;
+   if (doing_mixed) {
+     assert(old_generation()->available() >= old_generation()->free_unaffiliated_regions() * region_size_bytes,
+            "Unaffiliated available must be less than total available");
+ 
+     // We want this much memory to be unfragmented in order to reliably evacuate old.  This is conservative because we
+     // may not evacuate the entirety of unprocessed candidates in a single mixed evacuation.
+     size_t max_evac_need = (size_t)
+       (old_heuristics()->unprocessed_old_collection_candidates_live_memory() * ShenandoahOldEvacWaste);
+     size_t old_fragmented_available =
+       old_generation()->available() - old_generation()->free_unaffiliated_regions() * region_size_bytes;
+     reserve_for_mixed = max_evac_need + old_fragmented_available;
+     if (reserve_for_mixed > max_old_reserve) {
+       reserve_for_mixed = max_old_reserve;
+     }
+   } else {
+     reserve_for_mixed = 0;
+   }
+ 
+   size_t available_for_promotions = max_old_reserve - reserve_for_mixed;
+   if (doing_promotions) {
+     // We're only promoting and we have a maximum bound on the amount to be promoted
+     reserve_for_promo = (size_t) (promo_load * ShenandoahPromoEvacWaste);
+     if (reserve_for_promo > available_for_promotions) {
+       reserve_for_promo = available_for_promotions;
+     }
+   } else {
+     reserve_for_promo = 0;
+   }
+   old_reserve = reserve_for_mixed + reserve_for_promo;
+   assert(old_reserve <= max_old_reserve, "cannot reserve more than max for old evacuations");
+   size_t old_available = old_generation()->available() + old_cset_regions * region_size_bytes;
+   size_t young_available = young_generation()->available() + young_cset_regions * region_size_bytes;
+   size_t old_region_deficit = 0;
+   size_t old_region_surplus = 0;
+   if (old_available >= old_reserve) {
+     size_t old_excess = old_available - old_reserve;
+     size_t excess_regions = old_excess / region_size_bytes;
+     size_t unaffiliated_old_regions = old_generation()->free_unaffiliated_regions() + old_cset_regions;
+     size_t unaffiliated_old = unaffiliated_old_regions * region_size_bytes;
+     if (unaffiliated_old_regions < excess_regions) {
+       // We'll give only unaffiliated old to young, which is known to be less than the excess.
+       old_region_surplus = unaffiliated_old_regions;
+     } else {
+       // unaffiliated_old_regions > excess_regions, so we only give away the excess.
+       old_region_surplus = excess_regions;
+     }
+   } else {
+     // We need to request transfer from YOUNG.  Ignore that this will directly impact young_generation()->max_capacity(),
+     // indirectly impacting young_reserve and old_reserve.  These computations are conservative.
+     size_t old_need = old_reserve - old_available;
+     // Round up the number of regions needed from YOUNG
+     old_region_deficit = (old_need + region_size_bytes - 1) / region_size_bytes;
+   }
+   if (old_region_deficit > max_old_region_xfer) {
+     // If we're running short on young-gen memory, limit the xfer.  Old-gen collection activities will be curtailed
+     // if the budget is smaller than desired.
+     old_region_deficit = max_old_region_xfer;
+   }
+   set_old_region_surplus(old_region_surplus);
+   set_old_region_deficit(old_region_deficit);
+ }
+ 
+ // Called from stubs in JIT code or interpreter
  HeapWord* ShenandoahHeap::allocate_new_tlab(size_t min_size,
                                              size_t requested_size,
                                              size_t* actual_size) {
    ShenandoahAllocRequest req = ShenandoahAllocRequest::for_tlab(min_size, requested_size);
-   HeapWord* res = allocate_memory(req);
+   HeapWord* res = allocate_memory(req, false);
    if (res != nullptr) {
      *actual_size = req.actual_size();
    } else {
      *actual_size = 0;
    }

@@ -847,20 +1289,42 @@
  
  HeapWord* ShenandoahHeap::allocate_new_gclab(size_t min_size,
                                               size_t word_size,
                                               size_t* actual_size) {
    ShenandoahAllocRequest req = ShenandoahAllocRequest::for_gclab(min_size, word_size);
-   HeapWord* res = allocate_memory(req);
+   HeapWord* res = allocate_memory(req, false);
+   if (res != nullptr) {
+     *actual_size = req.actual_size();
+   } else {
+     *actual_size = 0;
+   }
+   return res;
+ }
+ 
+ HeapWord* ShenandoahHeap::allocate_new_plab(size_t min_size,
+                                             size_t word_size,
+                                             size_t* actual_size) {
+   // Align requested sizes to card sized multiples
+   size_t words_in_card = CardTable::card_size_in_words();
+   size_t align_mask = ~(words_in_card - 1);
+   min_size = (min_size + words_in_card - 1) & align_mask;
+   word_size = (word_size + words_in_card - 1) & align_mask;
+   ShenandoahAllocRequest req = ShenandoahAllocRequest::for_plab(min_size, word_size);
+   // Note that allocate_memory() sets a thread-local flag to prohibit further promotions by this thread
+   // if we are at risk of infringing on the old-gen evacuation budget.
+   HeapWord* res = allocate_memory(req, false);
    if (res != nullptr) {
      *actual_size = req.actual_size();
    } else {
      *actual_size = 0;
    }
    return res;
  }
  
- HeapWord* ShenandoahHeap::allocate_memory(ShenandoahAllocRequest& req) {
+ // is_promotion is true iff this allocation is known for sure to hold the result of young-gen evacuation
+ // to old-gen.  plab allocates are not known as such, since they may hold old-gen evacuations.
+ HeapWord* ShenandoahHeap::allocate_memory(ShenandoahAllocRequest& req, bool is_promotion) {
    intptr_t pacer_epoch = 0;
    bool in_new_region = false;
    HeapWord* result = nullptr;
  
    if (req.is_mutator_alloc()) {

@@ -868,11 +1332,11 @@
        pacer()->pace_for_alloc(req.size());
        pacer_epoch = pacer()->epoch();
      }
  
      if (!ShenandoahAllocFailureALot || !should_inject_alloc_failure()) {
-       result = allocate_memory_under_lock(req, in_new_region);
+       result = allocate_memory_under_lock(req, in_new_region, is_promotion);
      }
  
      // Allocation failed, block until control thread reacted, then retry allocation.
      //
      // It might happen that one of the threads requesting allocation would unblock

@@ -882,67 +1346,252 @@
      // one full GC has completed).
      size_t original_count = shenandoah_policy()->full_gc_count();
      while (result == nullptr
          && (_progress_last_gc.is_set() || original_count == shenandoah_policy()->full_gc_count())) {
        control_thread()->handle_alloc_failure(req);
-       result = allocate_memory_under_lock(req, in_new_region);
+       result = allocate_memory_under_lock(req, in_new_region, is_promotion);
      }
+ 
    } else {
      assert(req.is_gc_alloc(), "Can only accept GC allocs here");
-     result = allocate_memory_under_lock(req, in_new_region);
+     result = allocate_memory_under_lock(req, in_new_region, is_promotion);
      // Do not call handle_alloc_failure() here, because we cannot block.
      // The allocation failure would be handled by the LRB slowpath with handle_alloc_failure_evac().
    }
  
    if (in_new_region) {
      control_thread()->notify_heap_changed();
+     regulator_thread()->notify_heap_changed();
    }
  
+   if (result == nullptr) {
+     req.set_actual_size(0);
+   }
+ 
+   // This is called regardless of the outcome of the allocation to account
+   // for any waste created by retiring regions with this request.
+   increase_used(req);
+ 
    if (result != nullptr) {
      size_t requested = req.size();
      size_t actual = req.actual_size();
  
      assert (req.is_lab_alloc() || (requested == actual),
              "Only LAB allocations are elastic: %s, requested = " SIZE_FORMAT ", actual = " SIZE_FORMAT,
              ShenandoahAllocRequest::alloc_type_to_string(req.type()), requested, actual);
  
      if (req.is_mutator_alloc()) {
-       notify_mutator_alloc_words(actual, false);
- 
        // If we requested more than we were granted, give the rest back to pacer.
        // This only matters if we are in the same pacing epoch: do not try to unpace
        // over the budget for the other phase.
        if (ShenandoahPacing && (pacer_epoch > 0) && (requested > actual)) {
          pacer()->unpace_for_alloc(pacer_epoch, requested - actual);
        }
-     } else {
-       increase_used(actual*HeapWordSize);
      }
    }
  
    return result;
  }
  
- HeapWord* ShenandoahHeap::allocate_memory_under_lock(ShenandoahAllocRequest& req, bool& in_new_region) {
-   ShenandoahHeapLocker locker(lock());
-   return _free_set->allocate(req, in_new_region);
+ HeapWord* ShenandoahHeap::allocate_memory_under_lock(ShenandoahAllocRequest& req, bool& in_new_region, bool is_promotion) {
+   bool try_smaller_lab_size = false;
+   size_t smaller_lab_size;
+   {
+     // promotion_eligible pertains only to PLAB allocations, denoting that the PLAB is allowed to allocate for promotions.
+     bool promotion_eligible = false;
+     bool allow_allocation = true;
+     bool plab_alloc = false;
+     size_t requested_bytes = req.size() * HeapWordSize;
+     HeapWord* result = nullptr;
+     ShenandoahHeapLocker locker(lock());
+     Thread* thread = Thread::current();
+ 
+     if (mode()->is_generational()) {
+       if (req.affiliation() == YOUNG_GENERATION) {
+         if (req.is_mutator_alloc()) {
+           size_t young_words_available = young_generation()->available() / HeapWordSize;
+           if (ShenandoahElasticTLAB && req.is_lab_alloc() && (req.min_size() < young_words_available)) {
+             // Allow ourselves to try a smaller lab size even if requested_bytes <= young_available.  We may need a smaller
+             // lab size because young memory has become too fragmented.
+             try_smaller_lab_size = true;
+             smaller_lab_size = (young_words_available < req.size())? young_words_available: req.size();
+           } else if (req.size() > young_words_available) {
+             // Can't allocate because even min_size() is larger than remaining young_available
+             log_info(gc, ergo)("Unable to shrink %s alloc request of minimum size: " SIZE_FORMAT
+                                ", young words available: " SIZE_FORMAT, req.type_string(),
+                                HeapWordSize * (req.is_lab_alloc()? req.min_size(): req.size()), young_words_available);
+             return nullptr;
+           }
+         }
+       } else {                    // reg.affiliation() == OLD_GENERATION
+         assert(req.type() != ShenandoahAllocRequest::_alloc_gclab, "GCLAB pertains only to young-gen memory");
+         if (req.type() ==  ShenandoahAllocRequest::_alloc_plab) {
+           plab_alloc = true;
+           size_t promotion_avail = get_promoted_reserve();
+           size_t promotion_expended = get_promoted_expended();
+           if (promotion_expended + requested_bytes > promotion_avail) {
+             promotion_avail = 0;
+             if (get_old_evac_reserve() == 0) {
+               // There are no old-gen evacuations in this pass.  There's no value in creating a plab that cannot
+               // be used for promotions.
+               allow_allocation = false;
+             }
+           } else {
+             promotion_avail = promotion_avail - (promotion_expended + requested_bytes);
+             promotion_eligible = true;
+           }
+         } else if (is_promotion) {
+           // This is a shared alloc for promotion
+           size_t promotion_avail = get_promoted_reserve();
+           size_t promotion_expended = get_promoted_expended();
+           if (promotion_expended + requested_bytes > promotion_avail) {
+             promotion_avail = 0;
+           } else {
+             promotion_avail = promotion_avail - (promotion_expended + requested_bytes);
+           }
+           if (promotion_avail == 0) {
+             // We need to reserve the remaining memory for evacuation.  Reject this allocation.  The object will be
+             // evacuated to young-gen memory and promoted during a future GC pass.
+             return nullptr;
+           }
+           // Else, we'll allow the allocation to proceed.  (Since we hold heap lock, the tested condition remains true.)
+         } else {
+           // This is a shared allocation for evacuation.  Memory has already been reserved for this purpose.
+         }
+       }
+     } // This ends the is_generational() block
+ 
+     // First try the original request.  If TLAB request size is greater than available, allocate() will attempt to downsize
+     // request to fit within available memory.
+     result = (allow_allocation)? _free_set->allocate(req, in_new_region): nullptr;
+     if (result != nullptr) {
+       if (req.is_old()) {
+         ShenandoahThreadLocalData::reset_plab_promoted(thread);
+         if (req.is_gc_alloc()) {
+           bool disable_plab_promotions = false;
+           if (req.type() ==  ShenandoahAllocRequest::_alloc_plab) {
+             if (promotion_eligible) {
+               size_t actual_size = req.actual_size() * HeapWordSize;
+               // The actual size of the allocation may be larger than the requested bytes (due to alignment on card boundaries).
+               // If this puts us over our promotion budget, we need to disable future PLAB promotions for this thread.
+               if (get_promoted_expended() + actual_size <= get_promoted_reserve()) {
+                 // Assume the entirety of this PLAB will be used for promotion.  This prevents promotion from overreach.
+                 // When we retire this plab, we'll unexpend what we don't really use.
+                 ShenandoahThreadLocalData::enable_plab_promotions(thread);
+                 expend_promoted(actual_size);
+                 assert(get_promoted_expended() <= get_promoted_reserve(), "Do not expend more promotion than budgeted");
+                 ShenandoahThreadLocalData::set_plab_preallocated_promoted(thread, actual_size);
+               } else {
+                 disable_plab_promotions = true;
+               }
+             } else {
+               disable_plab_promotions = true;
+             }
+             if (disable_plab_promotions) {
+               // Disable promotions in this thread because entirety of this PLAB must be available to hold old-gen evacuations.
+               ShenandoahThreadLocalData::disable_plab_promotions(thread);
+               ShenandoahThreadLocalData::set_plab_preallocated_promoted(thread, 0);
+             }
+           } else if (is_promotion) {
+             // Shared promotion.  Assume size is requested_bytes.
+             expend_promoted(requested_bytes);
+             assert(get_promoted_expended() <= get_promoted_reserve(), "Do not expend more promotion than budgeted");
+           }
+         }
+ 
+         // Register the newly allocated object while we're holding the global lock since there's no synchronization
+         // built in to the implementation of register_object().  There are potential races when multiple independent
+         // threads are allocating objects, some of which might span the same card region.  For example, consider
+         // a card table's memory region within which three objects are being allocated by three different threads:
+         //
+         // objects being "concurrently" allocated:
+         //    [-----a------][-----b-----][--------------c------------------]
+         //            [---- card table memory range --------------]
+         //
+         // Before any objects are allocated, this card's memory range holds no objects.  Note that allocation of object a
+         //   wants to set the starts-object, first-start, and last-start attributes of the preceding card region.
+         //   allocation of object b wants to set the starts-object, first-start, and last-start attributes of this card region.
+         //   allocation of object c also wants to set the starts-object, first-start, and last-start attributes of this
+         //   card region.
+         //
+         // The thread allocating b and the thread allocating c can "race" in various ways, resulting in confusion, such as
+         // last-start representing object b while first-start represents object c.  This is why we need to require all
+         // register_object() invocations to be "mutually exclusive" with respect to each card's memory range.
+         ShenandoahHeap::heap()->card_scan()->register_object(result);
+       }
+     } else {
+       // The allocation failed.  If this was a plab allocation, We've already retired it and no longer have a plab.
+       if (req.is_old() && req.is_gc_alloc() && (req.type() == ShenandoahAllocRequest::_alloc_plab)) {
+         // We don't need to disable PLAB promotions because there is no PLAB.  We leave promotions enabled because
+         // this allows the surrounding infrastructure to retry alloc_plab_slow() with a smaller PLAB size.
+         ShenandoahThreadLocalData::set_plab_preallocated_promoted(thread, 0);
+       }
+     }
+     if ((result != nullptr) || !try_smaller_lab_size) {
+       return result;
+     }
+     // else, fall through to try_smaller_lab_size
+   } // This closes the block that holds the heap lock, releasing the lock.
+ 
+   // We failed to allocate the originally requested lab size.  Let's see if we can allocate a smaller lab size.
+   if (req.size() == smaller_lab_size) {
+     // If we were already trying to allocate min size, no value in attempting to repeat the same.  End the recursion.
+     return nullptr;
+   }
+ 
+   // We arrive here if the tlab allocation request can be resized to fit within young_available
+   assert((req.affiliation() == YOUNG_GENERATION) && req.is_lab_alloc() && req.is_mutator_alloc() &&
+          (smaller_lab_size < req.size()), "Only shrink allocation request size for TLAB allocations");
+ 
+   // By convention, ShenandoahAllocationRequest is primarily read-only.  The only mutable instance data is represented by
+   // actual_size(), which is overwritten with the size of the allocaion when the allocation request is satisfied.  We use a
+   // recursive call here rather than introducing new methods to mutate the existing ShenandoahAllocationRequest argument.
+   // Mutation of the existing object might result in astonishing results if calling contexts assume the content of immutable
+   // fields remain constant.  The original TLAB allocation request was for memory that exceeded the current capacity.  We'll
+   // attempt to allocate a smaller TLAB.  If this is successful, we'll update actual_size() of our incoming
+   // ShenandoahAllocRequest.  If the recursive request fails, we'll simply return nullptr.
+ 
+   // Note that we've relinquished the HeapLock and some other thread may perform additional allocation before our recursive
+   // call reacquires the lock.  If that happens, we will need another recursive call to further reduce the size of our request
+   // for each time another thread allocates young memory during the brief intervals that the heap lock is available to
+   // interfering threads.  We expect this interference to be rare.  The recursion bottoms out when young_available is
+   // smaller than req.min_size().  The inner-nested call to allocate_memory_under_lock() uses the same min_size() value
+   // as this call, but it uses a preferred size() that is smaller than our preferred size, and is no larger than what we most
+   // recently saw as the memory currently available within the young generation.
+ 
+   // TODO: At the expense of code clarity, we could rewrite this recursive solution to use iteration.  We need at most one
+   // extra instance of the ShenandoahAllocRequest, which we can re-initialize multiple times inside a loop, with one iteration
+   // of the loop required for each time the existing solution would recurse.  An iterative solution would be more efficient
+   // in CPU time and stack memory utilization.  The expectation is that it is very rare that we would recurse more than once
+   // so making this change is not currently seen as a high priority.
+ 
+   ShenandoahAllocRequest smaller_req = ShenandoahAllocRequest::for_tlab(req.min_size(), smaller_lab_size);
+ 
+   // Note that shrinking the preferred size gets us past the gatekeeper that checks whether there's available memory to
+   // satisfy the allocation request.  The reality is the actual TLAB size is likely to be even smaller, because it will
+   // depend on how much memory is available within mutator regions that are not yet fully used.
+   HeapWord* result = allocate_memory_under_lock(smaller_req, in_new_region, is_promotion);
+   if (result != nullptr) {
+     req.set_actual_size(smaller_req.actual_size());
+   }
+   return result;
  }
  
  HeapWord* ShenandoahHeap::mem_allocate(size_t size,
                                          bool*  gc_overhead_limit_was_exceeded) {
    ShenandoahAllocRequest req = ShenandoahAllocRequest::for_shared(size);
-   return allocate_memory(req);
+   return allocate_memory(req, false);
  }
  
  MetaWord* ShenandoahHeap::satisfy_failed_metadata_allocation(ClassLoaderData* loader_data,
                                                               size_t size,
                                                               Metaspace::MetadataType mdtype) {
    MetaWord* result;
  
    // Inform metaspace OOM to GC heuristics if class unloading is possible.
-   if (heuristics()->can_unload_classes()) {
-     ShenandoahHeuristics* h = heuristics();
+   ShenandoahHeuristics* h = global_generation()->heuristics();
+   if (h->can_unload_classes()) {
      h->record_metaspace_oom();
    }
  
    // Expand and retry allocation
    result = loader_data->metaspace_non_null()->expand_and_allocate(size, mdtype);

@@ -1017,26 +1666,123 @@
    void do_work() {
      ShenandoahConcurrentEvacuateRegionObjectClosure cl(_sh);
      ShenandoahHeapRegion* r;
      while ((r =_cs->claim_next()) != nullptr) {
        assert(r->has_live(), "Region " SIZE_FORMAT " should have been reclaimed early", r->index());
+ 
        _sh->marked_object_iterate(r, &cl);
  
        if (ShenandoahPacing) {
          _sh->pacer()->report_evac(r->used() >> LogHeapWordSize);
        }
+       if (_sh->check_cancelled_gc_and_yield(_concurrent)) {
+         break;
+       }
+     }
+   }
+ };
+ 
+ // Unlike ShenandoahEvacuationTask, this iterates over all regions rather than just the collection set.
+ // This is needed in order to promote humongous start regions if age() >= tenure threshold.
+ class ShenandoahGenerationalEvacuationTask : public WorkerTask {
+ private:
+   ShenandoahHeap* const _sh;
+   ShenandoahRegionIterator *_regions;
+   bool _concurrent;
+   uint _tenuring_threshold;
+ 
+ public:
+   ShenandoahGenerationalEvacuationTask(ShenandoahHeap* sh,
+                                        ShenandoahRegionIterator* iterator,
+                                        bool concurrent) :
+     WorkerTask("Shenandoah Evacuation"),
+     _sh(sh),
+     _regions(iterator),
+     _concurrent(concurrent),
+     _tenuring_threshold(0)
+   {
+     if (_sh->mode()->is_generational()) {
+       _tenuring_threshold = _sh->age_census()->tenuring_threshold();
+     }
+   }
+ 
+   void work(uint worker_id) {
+     if (_concurrent) {
+       ShenandoahConcurrentWorkerSession worker_session(worker_id);
+       ShenandoahSuspendibleThreadSetJoiner stsj(ShenandoahSuspendibleWorkers);
+       ShenandoahEvacOOMScope oom_evac_scope;
+       do_work();
+     } else {
+       ShenandoahParallelWorkerSession worker_session(worker_id);
+       ShenandoahEvacOOMScope oom_evac_scope;
+       do_work();
+     }
+   }
+ 
+ private:
+   void do_work() {
+     ShenandoahConcurrentEvacuateRegionObjectClosure cl(_sh);
+     ShenandoahHeapRegion* r;
+     ShenandoahMarkingContext* const ctx = ShenandoahHeap::heap()->marking_context();
+     size_t region_size_bytes = ShenandoahHeapRegion::region_size_bytes();
+     size_t old_garbage_threshold = (region_size_bytes * ShenandoahOldGarbageThreshold) / 100;
+     while ((r = _regions->next()) != nullptr) {
+       log_debug(gc)("GenerationalEvacuationTask do_work(), looking at %s region " SIZE_FORMAT ", (age: %d) [%s, %s, %s]",
+                     r->is_old()? "old": r->is_young()? "young": "free", r->index(), r->age(),
+                     r->is_active()? "active": "inactive",
+                     r->is_humongous()? (r->is_humongous_start()? "humongous_start": "humongous_continuation"): "regular",
+                     r->is_cset()? "cset": "not-cset");
+ 
+       if (r->is_cset()) {
+         assert(r->has_live(), "Region " SIZE_FORMAT " should have been reclaimed early", r->index());
+         _sh->marked_object_iterate(r, &cl);
+         if (ShenandoahPacing) {
+           _sh->pacer()->report_evac(r->used() >> LogHeapWordSize);
+         }
+       } else if (r->is_young() && r->is_active() && (r->age() >= _tenuring_threshold)) {
+         HeapWord* tams = ctx->top_at_mark_start(r);
+         if (r->is_humongous_start()) {
+           // We promote humongous_start regions along with their affiliated continuations during evacuation rather than
+           // doing this work during a safepoint.  We cannot put humongous regions into the collection set because that
+           // triggers the load-reference barrier (LRB) to copy on reference fetch.
+           r->promote_humongous();
+         } else if (r->is_regular() && (r->get_top_before_promote() != nullptr)) {
+           assert(r->garbage_before_padded_for_promote() < old_garbage_threshold,
+                  "Region " SIZE_FORMAT " has too much garbage for promotion", r->index());
+           assert(r->get_top_before_promote() == tams,
+                  "Region " SIZE_FORMAT " has been used for allocations before promotion", r->index());
+           // Likewise, we cannot put promote-in-place regions into the collection set because that would also trigger
+           // the LRB to copy on reference fetch.
+           r->promote_in_place();
+         }
+         // Aged humongous continuation regions are handled with their start region.  If an aged regular region has
+         // more garbage than ShenandoahOldGarbageTrheshold, we'll promote by evacuation.  If there is room for evacuation
+         // in this cycle, the region will be in the collection set.  If there is not room, the region will be promoted
+         // by evacuation in some future GC cycle.
+ 
+         // If an aged regular region has received allocations during the current cycle, we do not promote because the
+         // newly allocated objects do not have appropriate age; this region's age will be reset to zero at end of cycle.
+       }
+       // else, region is free, or OLD, or not in collection set, or humongous_continuation,
+       // or is young humongous_start that is too young to be promoted
  
        if (_sh->check_cancelled_gc_and_yield(_concurrent)) {
          break;
        }
      }
    }
  };
  
  void ShenandoahHeap::evacuate_collection_set(bool concurrent) {
-   ShenandoahEvacuationTask task(this, _collection_set, concurrent);
-   workers()->run_task(&task);
+   if (ShenandoahHeap::heap()->mode()->is_generational()) {
+     ShenandoahRegionIterator regions;
+     ShenandoahGenerationalEvacuationTask task(this, &regions, concurrent);
+     workers()->run_task(&task);
+   } else {
+     ShenandoahEvacuationTask task(this, _collection_set, concurrent);
+     workers()->run_task(&task);
+   }
  }
  
  void ShenandoahHeap::trash_cset_regions() {
    ShenandoahHeapLocker locker(lock());
  

@@ -1062,11 +1808,11 @@
    for (size_t i = 0; i < num_regions(); i++) {
      get_region(i)->print_on(st);
    }
  }
  
- void ShenandoahHeap::trash_humongous_region_at(ShenandoahHeapRegion* start) {
+ size_t ShenandoahHeap::trash_humongous_region_at(ShenandoahHeapRegion* start) {
    assert(start->is_humongous_start(), "reclaim regions starting with the first one");
  
    oop humongous_obj = cast_to_oop(start->bottom());
    size_t size = humongous_obj->size();
    size_t required_regions = ShenandoahHeapRegion::required_regions(size * HeapWordSize);

@@ -1082,19 +1828,24 @@
      assert(region->is_humongous(), "expect correct humongous start or continuation");
      assert(!region->is_cset(), "Humongous region should not be in collection set");
  
      region->make_trash_immediate();
    }
+   return required_regions;
  }
  
  class ShenandoahCheckCleanGCLABClosure : public ThreadClosure {
  public:
    ShenandoahCheckCleanGCLABClosure() {}
    void do_thread(Thread* thread) {
      PLAB* gclab = ShenandoahThreadLocalData::gclab(thread);
      assert(gclab != nullptr, "GCLAB should be initialized for %s", thread->name());
      assert(gclab->words_remaining() == 0, "GCLAB should not need retirement");
+ 
+     PLAB* plab = ShenandoahThreadLocalData::plab(thread);
+     assert(plab != nullptr, "PLAB should be initialized for %s", thread->name());
+     assert(plab->words_remaining() == 0, "PLAB should not need retirement");
    }
  };
  
  class ShenandoahRetireGCLABClosure : public ThreadClosure {
  private:

@@ -1106,10 +1857,21 @@
      assert(gclab != nullptr, "GCLAB should be initialized for %s", thread->name());
      gclab->retire();
      if (_resize && ShenandoahThreadLocalData::gclab_size(thread) > 0) {
        ShenandoahThreadLocalData::set_gclab_size(thread, 0);
      }
+ 
+     PLAB* plab = ShenandoahThreadLocalData::plab(thread);
+     assert(plab != nullptr, "PLAB should be initialized for %s", thread->name());
+ 
+     // There are two reasons to retire all plabs between old-gen evacuation passes.
+     //  1. We need to make the plab memory parseable by remembered-set scanning.
+     //  2. We need to establish a trustworthy UpdateWaterMark value within each old-gen heap region
+     ShenandoahHeap::heap()->retire_plab(plab, thread);
+     if (_resize && ShenandoahThreadLocalData::plab_size(thread) > 0) {
+       ShenandoahThreadLocalData::set_plab_size(thread, 0);
+     }
    }
  };
  
  void ShenandoahHeap::labs_make_parsable() {
    assert(UseTLAB, "Only call with UseTLAB");

@@ -1163,16 +1925,51 @@
    if (safepoint_workers() != nullptr) {
      safepoint_workers()->threads_do(&cl);
    }
  }
  
+ class ShenandoahTagGCLABClosure : public ThreadClosure {
+ public:
+   void do_thread(Thread* thread) {
+     PLAB* gclab = ShenandoahThreadLocalData::gclab(thread);
+     assert(gclab != nullptr, "GCLAB should be initialized for %s", thread->name());
+     if (gclab->words_remaining() > 0) {
+       ShenandoahHeapRegion* r = ShenandoahHeap::heap()->heap_region_containing(gclab->allocate(0));
+       r->set_young_lab_flag();
+     }
+   }
+ };
+ 
+ void ShenandoahHeap::set_young_lab_region_flags() {
+   if (!UseTLAB) {
+     return;
+   }
+   for (size_t i = 0; i < _num_regions; i++) {
+     _regions[i]->clear_young_lab_flags();
+   }
+   ShenandoahTagGCLABClosure cl;
+   workers()->threads_do(&cl);
+   for (JavaThreadIteratorWithHandle jtiwh; JavaThread *t = jtiwh.next(); ) {
+     cl.do_thread(t);
+     ThreadLocalAllocBuffer& tlab = t->tlab();
+     if (tlab.end() != nullptr) {
+       ShenandoahHeapRegion* r = heap_region_containing(tlab.start());
+       r->set_young_lab_flag();
+     }
+   }
+ }
+ 
  // Returns size in bytes
  size_t ShenandoahHeap::unsafe_max_tlab_alloc(Thread *thread) const {
    if (ShenandoahElasticTLAB) {
-     // With Elastic TLABs, return the max allowed size, and let the allocation path
-     // figure out the safe size for current allocation.
-     return ShenandoahHeapRegion::max_tlab_size_bytes();
+     if (mode()->is_generational()) {
+       return MIN2(ShenandoahHeapRegion::max_tlab_size_bytes(), young_generation()->available());
+     } else {
+       // With Elastic TLABs, return the max allowed size, and let the allocation path
+       // figure out the safe size for current allocation.
+       return ShenandoahHeapRegion::max_tlab_size_bytes();
+     }
    } else {
      return MIN2(_free_set->unsafe_peek_free(), ShenandoahHeapRegion::max_tlab_size_bytes());
    }
  }
  

@@ -1211,11 +2008,16 @@
      labs_make_parsable();
    }
  }
  
  void ShenandoahHeap::gc_threads_do(ThreadClosure* tcl) const {
+   if (_shenandoah_policy->is_at_shutdown()) {
+     return;
+   }
+ 
    tcl->do_thread(_control_thread);
+   tcl->do_thread(_regulator_thread);
    workers()->threads_do(tcl);
    if (_safepoint_workers != nullptr) {
      _safepoint_workers->threads_do(tcl);
    }
  }

@@ -1231,15 +2033,37 @@
      ls.cr();
      ls.cr();
  
      shenandoah_policy()->print_gc_stats(&ls);
  
+     ls.cr();
+ 
+     evac_tracker()->print_global_on(&ls);
+ 
      ls.cr();
      ls.cr();
    }
  }
  
+ void ShenandoahHeap::on_cycle_start(GCCause::Cause cause, ShenandoahGeneration* generation) {
+   set_gc_cause(cause);
+   set_gc_generation(generation);
+ 
+   shenandoah_policy()->record_cycle_start();
+   generation->heuristics()->record_cycle_start();
+ }
+ 
+ void ShenandoahHeap::on_cycle_end(ShenandoahGeneration* generation) {
+   generation->heuristics()->record_cycle_end();
+   if (mode()->is_generational() && (generation->is_global() || upgraded_to_full())) {
+     // If we just completed a GLOBAL GC, claim credit for completion of young-gen and old-gen GC as well
+     young_generation()->heuristics()->record_cycle_end();
+     old_generation()->heuristics()->record_cycle_end();
+   }
+   set_gc_cause(GCCause::_no_gc);
+ }
+ 
  void ShenandoahHeap::verify(VerifyOption vo) {
    if (ShenandoahSafepoint::is_at_shenandoah_safepoint()) {
      if (ShenandoahVerify) {
        verifier()->verify_generic(vo);
      } else {

@@ -1557,33 +2381,10 @@
    } else {
      heap_region_iterate(blk);
    }
  }
  
- class ShenandoahInitMarkUpdateRegionStateClosure : public ShenandoahHeapRegionClosure {
- private:
-   ShenandoahMarkingContext* const _ctx;
- public:
-   ShenandoahInitMarkUpdateRegionStateClosure() : _ctx(ShenandoahHeap::heap()->marking_context()) {}
- 
-   void heap_region_do(ShenandoahHeapRegion* r) {
-     assert(!r->has_live(), "Region " SIZE_FORMAT " should have no live data", r->index());
-     if (r->is_active()) {
-       // Check if region needs updating its TAMS. We have updated it already during concurrent
-       // reset, so it is very likely we don't need to do another write here.
-       if (_ctx->top_at_mark_start(r) != r->top()) {
-         _ctx->capture_top_at_mark_start(r);
-       }
-     } else {
-       assert(_ctx->top_at_mark_start(r) == r->top(),
-              "Region " SIZE_FORMAT " should already have correct TAMS", r->index());
-     }
-   }
- 
-   bool is_thread_safe() { return true; }
- };
- 
  class ShenandoahRendezvousClosure : public HandshakeClosure {
  public:
    inline ShenandoahRendezvousClosure() : HandshakeClosure("ShenandoahRendezvous") {}
    inline void do_thread(Thread* thread) {}
  };

@@ -1595,120 +2396,21 @@
  
  void ShenandoahHeap::recycle_trash() {
    free_set()->recycle_trash();
  }
  
- class ShenandoahResetUpdateRegionStateClosure : public ShenandoahHeapRegionClosure {
- private:
-   ShenandoahMarkingContext* const _ctx;
- public:
-   ShenandoahResetUpdateRegionStateClosure() : _ctx(ShenandoahHeap::heap()->marking_context()) {}
- 
-   void heap_region_do(ShenandoahHeapRegion* r) {
-     if (r->is_active()) {
-       // Reset live data and set TAMS optimistically. We would recheck these under the pause
-       // anyway to capture any updates that happened since now.
-       r->clear_live_data();
-       _ctx->capture_top_at_mark_start(r);
-     }
-   }
- 
-   bool is_thread_safe() { return true; }
- };
- 
- void ShenandoahHeap::prepare_gc() {
-   reset_mark_bitmap();
- 
-   ShenandoahResetUpdateRegionStateClosure cl;
-   parallel_heap_region_iterate(&cl);
- }
- 
- class ShenandoahFinalMarkUpdateRegionStateClosure : public ShenandoahHeapRegionClosure {
- private:
-   ShenandoahMarkingContext* const _ctx;
-   ShenandoahHeapLock* const _lock;
- 
- public:
-   ShenandoahFinalMarkUpdateRegionStateClosure() :
-     _ctx(ShenandoahHeap::heap()->complete_marking_context()), _lock(ShenandoahHeap::heap()->lock()) {}
- 
-   void heap_region_do(ShenandoahHeapRegion* r) {
-     if (r->is_active()) {
-       // All allocations past TAMS are implicitly live, adjust the region data.
-       // Bitmaps/TAMS are swapped at this point, so we need to poll complete bitmap.
-       HeapWord *tams = _ctx->top_at_mark_start(r);
-       HeapWord *top = r->top();
-       if (top > tams) {
-         r->increase_live_data_alloc_words(pointer_delta(top, tams));
-       }
- 
-       // We are about to select the collection set, make sure it knows about
-       // current pinning status. Also, this allows trashing more regions that
-       // now have their pinning status dropped.
-       if (r->is_pinned()) {
-         if (r->pin_count() == 0) {
-           ShenandoahHeapLocker locker(_lock);
-           r->make_unpinned();
-         }
-       } else {
-         if (r->pin_count() > 0) {
-           ShenandoahHeapLocker locker(_lock);
-           r->make_pinned();
-         }
-       }
- 
-       // Remember limit for updating refs. It's guaranteed that we get no
-       // from-space-refs written from here on.
-       r->set_update_watermark_at_safepoint(r->top());
-     } else {
-       assert(!r->has_live(), "Region " SIZE_FORMAT " should have no live data", r->index());
-       assert(_ctx->top_at_mark_start(r) == r->top(),
-              "Region " SIZE_FORMAT " should have correct TAMS", r->index());
-     }
-   }
- 
-   bool is_thread_safe() { return true; }
- };
- 
- void ShenandoahHeap::prepare_regions_and_collection_set(bool concurrent) {
-   assert(!is_full_gc_in_progress(), "Only for concurrent and degenerated GC");
-   {
-     ShenandoahGCPhase phase(concurrent ? ShenandoahPhaseTimings::final_update_region_states :
-                                          ShenandoahPhaseTimings::degen_gc_final_update_region_states);
-     ShenandoahFinalMarkUpdateRegionStateClosure cl;
-     parallel_heap_region_iterate(&cl);
- 
-     assert_pinned_region_status();
-   }
- 
-   {
-     ShenandoahGCPhase phase(concurrent ? ShenandoahPhaseTimings::choose_cset :
-                                          ShenandoahPhaseTimings::degen_gc_choose_cset);
-     ShenandoahHeapLocker locker(lock());
-     _collection_set->clear();
-     heuristics()->choose_collection_set(_collection_set);
-   }
- 
-   {
-     ShenandoahGCPhase phase(concurrent ? ShenandoahPhaseTimings::final_rebuild_freeset :
-                                          ShenandoahPhaseTimings::degen_gc_final_rebuild_freeset);
-     ShenandoahHeapLocker locker(lock());
-     _free_set->rebuild();
-   }
- }
- 
  void ShenandoahHeap::do_class_unloading() {
    _unloader.unload();
  }
  
  void ShenandoahHeap::stw_weak_refs(bool full_gc) {
    // Weak refs processing
    ShenandoahPhaseTimings::Phase phase = full_gc ? ShenandoahPhaseTimings::full_gc_weakrefs
                                                  : ShenandoahPhaseTimings::degen_gc_weakrefs;
    ShenandoahTimingsTracker t(phase);
    ShenandoahGCWorkerPhase worker_phase(phase);
-   ref_processor()->process_references(phase, workers(), false /* concurrent */);
+   active_generation()->ref_processor()->process_references(phase, workers(), false /* concurrent */);
  }
  
  void ShenandoahHeap::prepare_update_heap_references(bool concurrent) {
    assert(ShenandoahSafepoint::is_at_shenandoah_safepoint(), "must be at safepoint");
  

@@ -1735,14 +2437,70 @@
    assert(ShenandoahSafepoint::is_at_shenandoah_safepoint(), "Should really be Shenandoah safepoint");
    _gc_state.set_cond(mask, value);
    set_gc_state_all_threads(_gc_state.raw_value());
  }
  
- void ShenandoahHeap::set_concurrent_mark_in_progress(bool in_progress) {
-   assert(!has_forwarded_objects(), "Not expected before/after mark phase");
-   set_gc_state_mask(MARKING, in_progress);
-   ShenandoahBarrierSet::satb_mark_queue_set().set_active_all_threads(in_progress, !in_progress);
+ void ShenandoahHeap::set_evacuation_reserve_quantities(bool is_valid) {
+   _has_evacuation_reserve_quantities = is_valid;
+ }
+ 
+ void ShenandoahHeap::set_concurrent_young_mark_in_progress(bool in_progress) {
+   uint mask;
+   assert(!has_forwarded_objects(), "Young marking is not concurrent with evacuation");
+   if (!in_progress && is_concurrent_old_mark_in_progress()) {
+     assert(mode()->is_generational(), "Only generational GC has old marking");
+     assert(_gc_state.is_set(MARKING), "concurrent_old_marking_in_progress implies MARKING");
+     // If old-marking is in progress when we turn off YOUNG_MARKING, leave MARKING (and OLD_MARKING) on
+     mask = YOUNG_MARKING;
+   } else {
+     mask = MARKING | YOUNG_MARKING;
+   }
+   set_gc_state_mask(mask, in_progress);
+   manage_satb_barrier(in_progress);
+ }
+ 
+ void ShenandoahHeap::set_concurrent_old_mark_in_progress(bool in_progress) {
+ #ifdef ASSERT
+   // has_forwarded_objects() iff UPDATEREFS or EVACUATION
+   bool has_forwarded = has_forwarded_objects()? 1: 0;
+   bool updating_or_evacuating = _gc_state.is_set(UPDATEREFS | EVACUATION)? 1: 0;
+   assert (has_forwarded == updating_or_evacuating, "Has forwarded objects iff updating or evacuating");
+ #endif
+   if (!in_progress && is_concurrent_young_mark_in_progress()) {
+     // If young-marking is in progress when we turn off OLD_MARKING, leave MARKING (and YOUNG_MARKING) on
+     assert(_gc_state.is_set(MARKING), "concurrent_young_marking_in_progress implies MARKING");
+     set_gc_state_mask(OLD_MARKING, in_progress);
+   } else {
+     set_gc_state_mask(MARKING | OLD_MARKING, in_progress);
+   }
+   manage_satb_barrier(in_progress);
+ }
+ 
+ void ShenandoahHeap::set_prepare_for_old_mark_in_progress(bool in_progress) {
+   // Unlike other set-gc-state functions, this may happen outside safepoint.
+   // Is only set and queried by control thread, so no coherence issues.
+   _prepare_for_old_mark = in_progress;
+ }
+ 
+ void ShenandoahHeap::set_aging_cycle(bool in_progress) {
+   _is_aging_cycle.set_cond(in_progress);
+ }
+ 
+ void ShenandoahHeap::manage_satb_barrier(bool active) {
+   if (is_concurrent_mark_in_progress()) {
+     // Ignore request to deactivate barrier while concurrent mark is in progress.
+     // Do not attempt to re-activate the barrier if it is already active.
+     if (active && !ShenandoahBarrierSet::satb_mark_queue_set().is_active()) {
+       ShenandoahBarrierSet::satb_mark_queue_set().set_active_all_threads(active, !active);
+     }
+   } else {
+     // No concurrent marking is in progress so honor request to deactivate,
+     // but only if the barrier is already active.
+     if (!active && ShenandoahBarrierSet::satb_mark_queue_set().is_active()) {
+       ShenandoahBarrierSet::satb_mark_queue_set().set_active_all_threads(active, !active);
+     }
+   }
  }
  
  void ShenandoahHeap::set_evacuation_in_progress(bool in_progress) {
    assert(ShenandoahSafepoint::is_at_shenandoah_safepoint(), "Only call this at safepoint");
    set_gc_state_mask(EVACUATION, in_progress);

@@ -1771,37 +2529,52 @@
  bool ShenandoahHeap::try_cancel_gc() {
    jbyte prev = _cancelled_gc.cmpxchg(CANCELLED, CANCELLABLE);
    return prev == CANCELLABLE;
  }
  
+ void ShenandoahHeap::cancel_concurrent_mark() {
+   _young_generation->cancel_marking();
+   _old_generation->cancel_marking();
+   _global_generation->cancel_marking();
+ 
+   ShenandoahBarrierSet::satb_mark_queue_set().abandon_partial_marking();
+ }
+ 
  void ShenandoahHeap::cancel_gc(GCCause::Cause cause) {
    if (try_cancel_gc()) {
      FormatBuffer<> msg("Cancelling GC: %s", GCCause::to_string(cause));
      log_info(gc)("%s", msg.buffer());
      Events::log(Thread::current(), "%s", msg.buffer());
+     _cancel_requested_time = os::elapsedTime();
+     if (cause == GCCause::_shenandoah_upgrade_to_full_gc) {
+       _upgraded_to_full = true;
+     }
    }
  }
  
  uint ShenandoahHeap::max_workers() {
    return _max_workers;
  }
  
  void ShenandoahHeap::stop() {
    // The shutdown sequence should be able to terminate when GC is running.
  
-   // Step 0. Notify policy to disable event recording.
+   // Step 1. Notify policy to disable event recording and prevent visiting gc threads during shutdown
    _shenandoah_policy->record_shutdown();
  
-   // Step 1. Notify control thread that we are in shutdown.
+   // Step 2. Stop requesting collections.
+   regulator_thread()->stop();
+ 
+   // Step 3. Notify control thread that we are in shutdown.
    // Note that we cannot do that with stop(), because stop() is blocking and waits for the actual shutdown.
    // Doing stop() here would wait for the normal GC cycle to complete, never falling through to cancel below.
    control_thread()->prepare_for_graceful_shutdown();
  
-   // Step 2. Notify GC workers that we are cancelling GC.
+   // Step 4. Notify GC workers that we are cancelling GC.
    cancel_gc(GCCause::_shenandoah_stop_vm);
  
-   // Step 3. Wait until GC worker exits normally.
+   // Step 5. Wait until GC worker exits normally.
    control_thread()->stop();
  }
  
  void ShenandoahHeap::stw_unload_classes(bool full_gc) {
    if (!unload_classes()) return;

@@ -1891,24 +2664,21 @@
    ShenandoahHeap* heap = ShenandoahHeap::heap();
    assert(heap->collection_set() != nullptr, "Sanity");
    return (address) heap->collection_set()->biased_map_address();
  }
  
- address ShenandoahHeap::cancelled_gc_addr() {
-   return (address) ShenandoahHeap::heap()->_cancelled_gc.addr_of();
- }
- 
  address ShenandoahHeap::gc_state_addr() {
    return (address) ShenandoahHeap::heap()->_gc_state.addr_of();
  }
  
- size_t ShenandoahHeap::bytes_allocated_since_gc_start() const {
-   return Atomic::load(&_bytes_allocated_since_gc_start);
- }
- 
  void ShenandoahHeap::reset_bytes_allocated_since_gc_start() {
-   Atomic::store(&_bytes_allocated_since_gc_start, (size_t)0);
+   if (mode()->is_generational()) {
+     young_generation()->reset_bytes_allocated_since_gc_start();
+     old_generation()->reset_bytes_allocated_since_gc_start();
+   }
+ 
+   global_generation()->reset_bytes_allocated_since_gc_start();
  }
  
  void ShenandoahHeap::set_degenerated_gc_in_progress(bool in_progress) {
    _degenerated_gc_in_progress.set_cond(in_progress);
  }

@@ -1968,12 +2738,14 @@
  
  #ifdef ASSERT
  void ShenandoahHeap::assert_pinned_region_status() {
    for (size_t i = 0; i < num_regions(); i++) {
      ShenandoahHeapRegion* r = get_region(i);
-     assert((r->is_pinned() && r->pin_count() > 0) || (!r->is_pinned() && r->pin_count() == 0),
-            "Region " SIZE_FORMAT " pinning status is inconsistent", i);
+     if (active_generation()->contains(r)) {
+       assert((r->is_pinned() && r->pin_count() > 0) || (!r->is_pinned() && r->pin_count() == 0),
+              "Region " SIZE_FORMAT " pinning status is inconsistent", i);
+     }
    }
  }
  #endif
  
  ConcurrentGCTimer* ShenandoahHeap::gc_timer() const {

@@ -2029,75 +2801,268 @@
  template<bool CONCURRENT>
  class ShenandoahUpdateHeapRefsTask : public WorkerTask {
  private:
    ShenandoahHeap* _heap;
    ShenandoahRegionIterator* _regions;
+   ShenandoahRegionChunkIterator* _work_chunks;
+ 
  public:
-   ShenandoahUpdateHeapRefsTask(ShenandoahRegionIterator* regions) :
+   explicit ShenandoahUpdateHeapRefsTask(ShenandoahRegionIterator* regions,
+                                         ShenandoahRegionChunkIterator* work_chunks) :
      WorkerTask("Shenandoah Update References"),
      _heap(ShenandoahHeap::heap()),
-     _regions(regions) {
+     _regions(regions),
+     _work_chunks(work_chunks)
+   {
    }
  
    void work(uint worker_id) {
      if (CONCURRENT) {
        ShenandoahConcurrentWorkerSession worker_session(worker_id);
        ShenandoahSuspendibleThreadSetJoiner stsj(ShenandoahSuspendibleWorkers);
-       do_work<ShenandoahConcUpdateRefsClosure>();
+       do_work<ShenandoahConcUpdateRefsClosure>(worker_id);
      } else {
        ShenandoahParallelWorkerSession worker_session(worker_id);
-       do_work<ShenandoahSTWUpdateRefsClosure>();
+       do_work<ShenandoahSTWUpdateRefsClosure>(worker_id);
      }
    }
  
  private:
    template<class T>
-   void do_work() {
+   void do_work(uint worker_id) {
      T cl;
+     if (CONCURRENT && (worker_id == 0)) {
+       // We ask the first worker to replenish the Mutator free set by moving regions previously reserved to hold the
+       // results of evacuation.  These reserves are no longer necessary because evacuation has completed.
+       size_t cset_regions = _heap->collection_set()->count();
+       // We cannot transfer any more regions than will be reclaimed when the existing collection set is recycled, because
+       // we need the reclaimed collection set regions to replenish the collector reserves
+       _heap->free_set()->move_collector_sets_to_mutator(cset_regions);
+     }
+     // If !CONCURRENT, there's no value in expanding Mutator free set
+ 
      ShenandoahHeapRegion* r = _regions->next();
-     ShenandoahMarkingContext* const ctx = _heap->complete_marking_context();
+     // We update references for global, old, and young collections.
+     assert(_heap->active_generation()->is_mark_complete(), "Expected complete marking");
+     ShenandoahMarkingContext* const ctx = _heap->marking_context();
+     bool is_mixed = _heap->collection_set()->has_old_regions();
      while (r != nullptr) {
        HeapWord* update_watermark = r->get_update_watermark();
        assert (update_watermark >= r->bottom(), "sanity");
+ 
+       log_debug(gc)("ShenandoahUpdateHeapRefsTask::do_work(%u) looking at region " SIZE_FORMAT, worker_id, r->index());
+       bool region_progress = false;
        if (r->is_active() && !r->is_cset()) {
-         _heap->marked_object_oop_iterate(r, &cl, update_watermark);
+         if (!_heap->mode()->is_generational() || r->is_young()) {
+           _heap->marked_object_oop_iterate(r, &cl, update_watermark);
+           region_progress = true;
+         } else if (r->is_old()) {
+           if (_heap->active_generation()->is_global()) {
+             // Note that GLOBAL collection is not as effectively balanced as young and mixed cycles.  This is because
+             // concurrent GC threads are parceled out entire heap regions of work at a time and there
+             // is no "catchup phase" consisting of remembered set scanning, during which parcels of work are smaller
+             // and more easily distributed more fairly across threads.
+ 
+             // TODO: Consider an improvement to load balance GLOBAL GC.
+             _heap->marked_object_oop_iterate(r, &cl, update_watermark);
+             region_progress = true;
+           }
+           // Otherwise, this is an old region in a young or mixed cycle.  Process it during a second phase, below.
+           // Don't bother to report pacing progress in this case.
+         } else {
+           // Because updating of references runs concurrently, it is possible that a FREE inactive region transitions
+           // to a non-free active region while this loop is executing.  Whenever this happens, the changing of a region's
+           // active status may propagate at a different speed than the changing of the region's affiliation.
+ 
+           // When we reach this control point, it is because a race has allowed a region's is_active() status to be seen
+           // by this thread before the region's affiliation() is seen by this thread.
+ 
+           // It's ok for this race to occur because the newly transformed region does not have any references to be
+           // updated.
+ 
+           assert(r->get_update_watermark() == r->bottom(),
+                  "%s Region " SIZE_FORMAT " is_active but not recognized as YOUNG or OLD so must be newly transitioned from FREE",
+                  r->affiliation_name(), r->index());
+         }
        }
-       if (ShenandoahPacing) {
+       if (region_progress && ShenandoahPacing) {
          _heap->pacer()->report_updaterefs(pointer_delta(update_watermark, r->bottom()));
        }
        if (_heap->check_cancelled_gc_and_yield(CONCURRENT)) {
          return;
        }
        r = _regions->next();
      }
+ 
+     if (_heap->mode()->is_generational() && !_heap->active_generation()->is_global()) {
+       // Since this is generational and not GLOBAL, we have to process the remembered set.  There's no remembered
+       // set processing if not in generational mode or if GLOBAL mode.
+ 
+       // After this thread has exhausted its traditional update-refs work, it continues with updating refs within remembered set.
+       // The remembered set workload is better balanced between threads, so threads that are "behind" can catch up with other
+       // threads during this phase, allowing all threads to work more effectively in parallel.
+       struct ShenandoahRegionChunk assignment;
+       RememberedScanner* scanner = _heap->card_scan();
+ 
+       while (!_heap->check_cancelled_gc_and_yield(CONCURRENT) && _work_chunks->next(&assignment)) {
+         // Keep grabbing next work chunk to process until finished, or asked to yield
+         ShenandoahHeapRegion* r = assignment._r;
+         if (r->is_active() && !r->is_cset() && r->is_old()) {
+           HeapWord* start_of_range = r->bottom() + assignment._chunk_offset;
+           HeapWord* end_of_range = r->get_update_watermark();
+           if (end_of_range > start_of_range + assignment._chunk_size) {
+             end_of_range = start_of_range + assignment._chunk_size;
+           }
+ 
+           // Old region in a young cycle or mixed cycle.
+           if (is_mixed) {
+             // TODO: For mixed evac, consider building an old-gen remembered set that allows restricted updating
+             // within old-gen HeapRegions.  This remembered set can be constructed by old-gen concurrent marking
+             // and augmented by card marking.  For example, old-gen concurrent marking can remember for each old-gen
+             // card which other old-gen regions it refers to: none, one-other specifically, multiple-other non-specific.
+             // Update-references when _mixed_evac processess each old-gen memory range that has a traditional DIRTY
+             // card or if the "old-gen remembered set" indicates that this card holds pointers specifically to an
+             // old-gen region in the most recent collection set, or if this card holds pointers to other non-specific
+             // old-gen heap regions.
+ 
+             if (r->is_humongous()) {
+               if (start_of_range < end_of_range) {
+                 // Need to examine both dirty and clean cards during mixed evac.
+                 r->oop_iterate_humongous_slice(&cl, false, start_of_range, assignment._chunk_size, true);
+               }
+             } else {
+               // Since this is mixed evacuation, old regions that are candidates for collection have not been coalesced
+               // and filled.  Use mark bits to find objects that need to be updated.
+               //
+               // Future TODO: establish a second remembered set to identify which old-gen regions point to other old-gen
+               // regions which are in the collection set for a particular mixed evacuation.
+               if (start_of_range < end_of_range) {
+                 HeapWord* p = nullptr;
+                 size_t card_index = scanner->card_index_for_addr(start_of_range);
+                 // In case last object in my range spans boundary of my chunk, I may need to scan all the way to top()
+                 ShenandoahObjectToOopBoundedClosure<T> objs(&cl, start_of_range, r->top());
+ 
+                 // Any object that begins in a previous range is part of a different scanning assignment.  Any object that
+                 // starts after end_of_range is also not my responsibility.  (Either allocated during evacuation, so does
+                 // not hold pointers to from-space, or is beyond the range of my assigned work chunk.)
+ 
+                 // Find the first object that begins in my range, if there is one.
+                 p = start_of_range;
+                 oop obj = cast_to_oop(p);
+                 HeapWord* tams = ctx->top_at_mark_start(r);
+                 if (p >= tams) {
+                   // We cannot use ctx->is_marked(obj) to test whether an object begins at this address.  Instead,
+                   // we need to use the remembered set crossing map to advance p to the first object that starts
+                   // within the enclosing card.
+ 
+                   while (true) {
+                     HeapWord* first_object = scanner->first_object_in_card(card_index);
+                     if (first_object != nullptr) {
+                       p = first_object;
+                       break;
+                     } else if (scanner->addr_for_card_index(card_index + 1) < end_of_range) {
+                       card_index++;
+                     } else {
+                       // Force the loop that follows to immediately terminate.
+                       p = end_of_range;
+                       break;
+                     }
+                   }
+                   obj = cast_to_oop(p);
+                   // Note: p may be >= end_of_range
+                 } else if (!ctx->is_marked(obj)) {
+                   p = ctx->get_next_marked_addr(p, tams);
+                   obj = cast_to_oop(p);
+                   // If there are no more marked objects before tams, this returns tams.
+                   // Note that tams is either >= end_of_range, or tams is the start of an object that is marked.
+                 }
+                 while (p < end_of_range) {
+                   // p is known to point to the beginning of marked object obj
+                   objs.do_object(obj);
+                   HeapWord* prev_p = p;
+                   p += obj->size();
+                   if (p < tams) {
+                     p = ctx->get_next_marked_addr(p, tams);
+                     // If there are no more marked objects before tams, this returns tams.  Note that tams is
+                     // either >= end_of_range, or tams is the start of an object that is marked.
+                   }
+                   assert(p != prev_p, "Lack of forward progress");
+                   obj = cast_to_oop(p);
+                 }
+               }
+             }
+           } else {
+             // This is a young evac..
+             if (start_of_range < end_of_range) {
+               size_t cluster_size =
+                 CardTable::card_size_in_words() * ShenandoahCardCluster<ShenandoahDirectCardMarkRememberedSet>::CardsPerCluster;
+               size_t clusters = assignment._chunk_size / cluster_size;
+               assert(clusters * cluster_size == assignment._chunk_size, "Chunk assignment must align on cluster boundaries");
+               scanner->process_region_slice(r, assignment._chunk_offset, clusters, end_of_range, &cl, true, worker_id);
+             }
+           }
+           if (ShenandoahPacing && (start_of_range < end_of_range)) {
+             _heap->pacer()->report_updaterefs(pointer_delta(end_of_range, start_of_range));
+           }
+         }
+       }
+     }
    }
  };
  
  void ShenandoahHeap::update_heap_references(bool concurrent) {
    assert(!is_full_gc_in_progress(), "Only for concurrent and degenerated GC");
+   uint nworkers = workers()->active_workers();
+   ShenandoahRegionChunkIterator work_list(nworkers);
  
    if (concurrent) {
-     ShenandoahUpdateHeapRefsTask<true> task(&_update_refs_iterator);
+     ShenandoahUpdateHeapRefsTask<true> task(&_update_refs_iterator, &work_list);
      workers()->run_task(&task);
    } else {
-     ShenandoahUpdateHeapRefsTask<false> task(&_update_refs_iterator);
+     ShenandoahUpdateHeapRefsTask<false> task(&_update_refs_iterator, &work_list);
      workers()->run_task(&task);
    }
+   if (ShenandoahEnableCardStats && card_scan()!=nullptr) { // generational check proxy
+     card_scan()->log_card_stats(nworkers, CARD_STAT_UPDATE_REFS);
+   }
  }
  
- 
  class ShenandoahFinalUpdateRefsUpdateRegionStateClosure : public ShenandoahHeapRegionClosure {
  private:
+   ShenandoahMarkingContext* _ctx;
    ShenandoahHeapLock* const _lock;
+   bool _is_generational;
  
  public:
-   ShenandoahFinalUpdateRefsUpdateRegionStateClosure() : _lock(ShenandoahHeap::heap()->lock()) {}
+   ShenandoahFinalUpdateRefsUpdateRegionStateClosure(
+     ShenandoahMarkingContext* ctx) : _ctx(ctx), _lock(ShenandoahHeap::heap()->lock()),
+                                      _is_generational(ShenandoahHeap::heap()->mode()->is_generational()) { }
  
    void heap_region_do(ShenandoahHeapRegion* r) {
+ 
+     // Maintenance of region age must follow evacuation in order to account for evacuation allocations within survivor
+     // regions.  We consult region age during the subsequent evacuation to determine whether certain objects need to
+     // be promoted.
+     if (_is_generational && r->is_young() && r->is_active()) {
+       HeapWord *tams = _ctx->top_at_mark_start(r);
+       HeapWord *top = r->top();
+ 
+       // Allocations move the watermark when top moves.  However compacting
+       // objects will sometimes lower top beneath the watermark, after which,
+       // attempts to read the watermark will assert out (watermark should not be
+       // higher than top).
+       if (top > tams) {
+         // There have been allocations in this region since the start of the cycle.
+         // Any objects new to this region must not assimilate elevated age.
+         r->reset_age();
+       } else if (ShenandoahHeap::heap()->is_aging_cycle()) {
+         r->increment_age();
+       }
+     }
+ 
      // Drop unnecessary "pinned" state from regions that does not have CP marks
      // anymore, as this would allow trashing them.
- 
      if (r->is_active()) {
        if (r->is_pinned()) {
          if (r->pin_count() == 0) {
            ShenandoahHeapLocker locker(_lock);
            r->make_unpinned();

@@ -2120,11 +3085,11 @@
  
    {
      ShenandoahGCPhase phase(concurrent ?
                              ShenandoahPhaseTimings::final_update_refs_update_region_states :
                              ShenandoahPhaseTimings::degen_gc_final_update_refs_update_region_states);
-     ShenandoahFinalUpdateRefsUpdateRegionStateClosure cl;
+     ShenandoahFinalUpdateRefsUpdateRegionStateClosure cl (active_generation()->complete_marking_context());
      parallel_heap_region_iterate(&cl);
  
      assert_pinned_region_status();
    }
  

@@ -2135,16 +3100,64 @@
      trash_cset_regions();
    }
  }
  
  void ShenandoahHeap::rebuild_free_set(bool concurrent) {
-   {
-     ShenandoahGCPhase phase(concurrent ?
-                             ShenandoahPhaseTimings::final_update_refs_rebuild_freeset :
-                             ShenandoahPhaseTimings::degen_gc_final_update_refs_rebuild_freeset);
-     ShenandoahHeapLocker locker(lock());
-     _free_set->rebuild();
+   ShenandoahGCPhase phase(concurrent ?
+                           ShenandoahPhaseTimings::final_update_refs_rebuild_freeset :
+                           ShenandoahPhaseTimings::degen_gc_final_update_refs_rebuild_freeset);
+   size_t region_size_bytes = ShenandoahHeapRegion::region_size_bytes();
+   ShenandoahHeapLocker locker(lock());
+   size_t young_cset_regions, old_cset_regions;
+   _free_set->prepare_to_rebuild(young_cset_regions, old_cset_regions);
+ 
+   if (mode()->is_generational()) {
+     assert(verify_generation_usage(true, old_generation()->used_regions(),
+                                    old_generation()->used(), old_generation()->get_humongous_waste(),
+                                    true, young_generation()->used_regions(),
+                                    young_generation()->used(), young_generation()->get_humongous_waste()),
+            "Generation accounts are inaccurate");
+ 
+     // The computation of bytes_of_allocation_runway_before_gc_trigger is quite conservative so consider all of this
+     // available for transfer to old. Note that transfer of humongous regions does not impact available.
+     size_t allocation_runway = young_heuristics()->bytes_of_allocation_runway_before_gc_trigger(young_cset_regions);
+     adjust_generation_sizes_for_next_cycle(allocation_runway, young_cset_regions, old_cset_regions);
+ 
+     // Total old_available may have been expanded to hold anticipated promotions.  We trigger if the fragmented available
+     // memory represents more than 16 regions worth of data.  Note that fragmentation may increase when we promote regular
+     // regions in place when many of these regular regions have an abundant amount of available memory within them.  Fragmentation
+     // will decrease as promote-by-copy consumes the available memory within these partially consumed regions.
+     //
+     // We consider old-gen to have excessive fragmentation if more than 12.5% of old-gen is free memory that resides
+     // within partially consumed regions of memory.
+   }
+   // Rebuild free set based on adjusted generation sizes.
+   _free_set->rebuild(young_cset_regions, old_cset_regions);
+ 
+   if (mode()->is_generational()) {
+     size_t old_available = old_generation()->available();
+     size_t old_unaffiliated_available = old_generation()->free_unaffiliated_regions() * region_size_bytes;
+     size_t old_fragmented_available;
+     assert(old_available >= old_unaffiliated_available, "unaffiliated available is a subset of total available");
+     old_fragmented_available = old_available - old_unaffiliated_available;
+ 
+     size_t old_capacity = old_generation()->max_capacity();
+     size_t heap_capacity = capacity();
+     if ((old_capacity > heap_capacity / 8) && (old_fragmented_available > old_capacity / 8)) {
+       old_heuristics()->trigger_old_is_fragmented();
+     }
+ 
+     size_t old_used = old_generation()->used() + old_generation()->get_humongous_waste();
+     size_t trigger_threshold = old_generation()->usage_trigger_threshold();
+     // Detects unsigned arithmetic underflow
+     assert(old_used <= capacity(),
+            "Old used (" SIZE_FORMAT ", " SIZE_FORMAT") must not be more than heap capacity (" SIZE_FORMAT ")",
+            old_generation()->used(), old_generation()->get_humongous_waste(), capacity());
+ 
+     if (old_used > trigger_threshold) {
+       old_heuristics()->trigger_old_has_grown();
+     }
    }
  }
  
  void ShenandoahHeap::print_extended_on(outputStream *st) const {
    print_on(st);

@@ -2255,13 +3268,22 @@
  bool ShenandoahHeap::should_inject_alloc_failure() {
    return _inject_alloc_failure.is_set() && _inject_alloc_failure.try_unset();
  }
  
  void ShenandoahHeap::initialize_serviceability() {
-   _memory_pool = new ShenandoahMemoryPool(this);
-   _cycle_memory_manager.add_pool(_memory_pool);
-   _stw_memory_manager.add_pool(_memory_pool);
+   if (mode()->is_generational()) {
+     _young_gen_memory_pool = new ShenandoahYoungGenMemoryPool(this);
+     _old_gen_memory_pool = new ShenandoahOldGenMemoryPool(this);
+     _cycle_memory_manager.add_pool(_young_gen_memory_pool);
+     _cycle_memory_manager.add_pool(_old_gen_memory_pool);
+     _stw_memory_manager.add_pool(_young_gen_memory_pool);
+     _stw_memory_manager.add_pool(_old_gen_memory_pool);
+   } else {
+     _memory_pool = new ShenandoahMemoryPool(this);
+     _cycle_memory_manager.add_pool(_memory_pool);
+     _stw_memory_manager.add_pool(_memory_pool);
+   }
  }
  
  GrowableArray<GCMemoryManager*> ShenandoahHeap::memory_managers() {
    GrowableArray<GCMemoryManager*> memory_managers(2);
    memory_managers.append(&_cycle_memory_manager);

@@ -2269,16 +3291,21 @@
    return memory_managers;
  }
  
  GrowableArray<MemoryPool*> ShenandoahHeap::memory_pools() {
    GrowableArray<MemoryPool*> memory_pools(1);
-   memory_pools.append(_memory_pool);
+   if (mode()->is_generational()) {
+     memory_pools.append(_young_gen_memory_pool);
+     memory_pools.append(_old_gen_memory_pool);
+   } else {
+     memory_pools.append(_memory_pool);
+   }
    return memory_pools;
  }
  
  MemoryUsage ShenandoahHeap::memory_usage() {
-   return _memory_pool->get_memory_usage();
+   return MemoryUsage(_initial_size, used(), committed(), max_capacity());
  }
  
  ShenandoahRegionIterator::ShenandoahRegionIterator() :
    _heap(ShenandoahHeap::heap()),
    _index(0) {}

@@ -2312,10 +3339,11 @@
  
  void ShenandoahHeap::flush_liveness_cache(uint worker_id) {
    assert(worker_id < _max_workers, "sanity");
    assert(_liveness_cache != nullptr, "sanity");
    ShenandoahLiveData* ld = _liveness_cache[worker_id];
+ 
    for (uint i = 0; i < num_regions(); i++) {
      ShenandoahLiveData live = ld[i];
      if (live > 0) {
        ShenandoahHeapRegion* r = get_region(i);
        r->increase_live_data_gc_words(live);

@@ -2339,5 +3367,108 @@
      return true;
    }
  
    return false;
  }
+ 
+ void ShenandoahHeap::transfer_old_pointers_from_satb() {
+   _old_generation->transfer_pointers_from_satb();
+ }
+ 
+ template<>
+ void ShenandoahGenerationRegionClosure<YOUNG>::heap_region_do(ShenandoahHeapRegion* region) {
+   // Visit young and free regions
+   if (!region->is_old()) {
+     _cl->heap_region_do(region);
+   }
+ }
+ 
+ template<>
+ void ShenandoahGenerationRegionClosure<OLD>::heap_region_do(ShenandoahHeapRegion* region) {
+   // Visit old and free regions
+   if (!region->is_young()) {
+     _cl->heap_region_do(region);
+   }
+ }
+ 
+ template<>
+ void ShenandoahGenerationRegionClosure<GLOBAL_GEN>::heap_region_do(ShenandoahHeapRegion* region) {
+   _cl->heap_region_do(region);
+ }
+ 
+ template<>
+ void ShenandoahGenerationRegionClosure<GLOBAL_NON_GEN>::heap_region_do(ShenandoahHeapRegion* region) {
+   _cl->heap_region_do(region);
+ }
+ 
+ bool ShenandoahHeap::verify_generation_usage(bool verify_old, size_t old_regions, size_t old_bytes, size_t old_waste,
+                                              bool verify_young, size_t young_regions, size_t young_bytes, size_t young_waste) {
+   size_t tally_old_regions = 0;
+   size_t tally_old_bytes = 0;
+   size_t tally_old_waste = 0;
+   size_t tally_young_regions = 0;
+   size_t tally_young_bytes = 0;
+   size_t tally_young_waste = 0;
+ 
+   shenandoah_assert_heaplocked_or_safepoint();
+   for (size_t i = 0; i < num_regions(); i++) {
+     ShenandoahHeapRegion* r = get_region(i);
+     if (r->is_old()) {
+       tally_old_regions++;
+       tally_old_bytes += r->used();
+       if (r->is_humongous()) {
+         ShenandoahHeapRegion* start = r->humongous_start_region();
+         HeapWord* obj_addr = start->bottom();
+         oop obj = cast_to_oop(obj_addr);
+         size_t word_size = obj->size();
+         HeapWord* end_addr = obj_addr + word_size;
+         if (end_addr <= r->end()) {
+           tally_old_waste += (r->end() - end_addr) * HeapWordSize;
+         }
+       }
+     } else if (r->is_young()) {
+       tally_young_regions++;
+       tally_young_bytes += r->used();
+       if (r->is_humongous()) {
+         ShenandoahHeapRegion* start = r->humongous_start_region();
+         HeapWord* obj_addr = start->bottom();
+         oop obj = cast_to_oop(obj_addr);
+         size_t word_size = obj->size();
+         HeapWord* end_addr = obj_addr + word_size;
+         if (end_addr <= r->end()) {
+           tally_young_waste += (r->end() - end_addr) * HeapWordSize;
+         }
+       }
+     }
+   }
+   if (verify_young &&
+       ((young_regions != tally_young_regions) || (young_bytes != tally_young_bytes) || (young_waste != tally_young_waste))) {
+     return false;
+   } else if (verify_old &&
+              ((old_regions != tally_old_regions) || (old_bytes != tally_old_bytes) || (old_waste != tally_old_waste))) {
+     return false;
+   } else {
+     return true;
+   }
+ }
+ 
+ ShenandoahGeneration* ShenandoahHeap::generation_for(ShenandoahAffiliation affiliation) const {
+   if (!mode()->is_generational()) {
+     return global_generation();
+   } else if (affiliation == YOUNG_GENERATION) {
+     return young_generation();
+   } else if (affiliation == OLD_GENERATION) {
+     return old_generation();
+   }
+ 
+   ShouldNotReachHere();
+   return nullptr;
+ }
+ 
+ void ShenandoahHeap::log_heap_status(const char* msg) const {
+   if (mode()->is_generational()) {
+     young_generation()->log_status(msg);
+     old_generation()->log_status(msg);
+   } else {
+     global_generation()->log_status(msg);
+   }
+ }
< prev index next >